# Agents' next frontier: agent-to-agent and network effects — Jean-Denis Greze, Town

Jean-Denis Greze, Town

[Talk page](https://ai.engineer/talks/REascnFlq_8-agents-next-frontier-agent-agent-network-effects) · [Video](https://www.youtube.com/watch?v=REascnFlq_8)

## Agents’ next frontier: agent-to-agent and network effects

Jean-Denis Greze reframes agent collaboration as a search problem: how can the right information reach the next model call when it belongs to someone else? Five approaches move the privacy decision to different points in that process.

### The useful unit is the call with the right context

[0:12](https://www.youtube.com/watch?v=REascnFlq_8&t=12s)

People get work done with other people. Assistants should benefit from the same connections: another person’s agent may hold the information needed to complete your task. Jean-Denis Greze, introducing himself as Town’s CTO after seven years as CTO at Plaid, starts from that opportunity and immediately questions its framing. “Agent to agent” tells us who communicates. It does less to explain what makes the result useful.

The alternative framing is search. Immediately before a model returns an answer or calls a tool, its context window needs the information required for that decision. A more capable model can do more with that information, but the surrounding system still has to find it and bring it into context. Collaboration matters when it improves that process. [1:13](https://www.youtube.com/watch?v=REascnFlq_8&t=73s)

Greze sketches a progression from humans manually filling the context window, to retrieval-augmented generation using a search tool, to agentic search using multiple tools to explore available content. Each changes how information arrives. The engineering target remains the consequential model call: does it have the content needed to choose the right action?

### The impossible agent that can see everything

[2:13](https://www.youtube.com/watch?v=REascnFlq_8&t=133s)

Imagine a single agent with access to everyone’s email, every company’s records, and every government’s information. In Greze’s thought experiment, it can assemble all the relevant context before answering. This gives multi-agent systems a comparison point: how closely can they approach the information available to that agent? The optimal outcome is an idealization within this model of the problem, rather than a demonstrated guarantee that complete access makes an LLM correct.

Even an infinite context window would not make this agent acceptable. People do not grant strangers continuous access to their inboxes. Privacy and security restrict what can enter the search process before context length becomes relevant. Greze invokes the Coase theorem as an analogy: removing information barriers and transaction costs can improve negotiation outcomes, while real-world constraints prevent the frictionless ideal. Here, the constraint is permission to use information.

The five approaches that follow place that permission decision in different places. Some restrict the searchable data in advance. Others permit broad internal search and control what leaves it. Their usefulness depends both on the information they recover and on how much human work they require as models improve.

### Restrict access, or restrict what a tool returns

[4:14](https://www.youtube.com/watch?v=REascnFlq_8&t=254s)

The first strategy gives an agent broad access inside an existing trust boundary. An HR team agent might read the HR systems available to an employee, perhaps using the access level of the team’s most junior member. Everyone on the team can ask it questions. This fits a familiar SaaS security model, which helps explain its appeal to IT and security teams.

Its limitation is structural. Humans still choose the accessible data, and information outside that selection remains outside it. Better models may answer better within the permitted material, but they do not automatically dissolve the divisions between teams. Greze’s two tests are whether the system naturally needs fewer humans over time and whether model improvements expand what the approach can accomplish. A manually defined team silo fares poorly on both.

The second strategy builds tools that trade some information for a useful capability. Consider the question that returns throughout the talk: does anyone at the company know someone on Acme Corp’s finance team? Unrestricted search would read everyone’s Gmail, identify exchanges with Acme employees, inspect profiles, and find a connection to the CFO. That would also expose far more correspondence than the requester needs.

A narrower tool can accept a company domain and a target role, search the mail internally, and return ranked relationship-strength scores. The requesting agent receives a useful result without receiving the underlying emails. It then messages Bob on Slack about his connection to Jane, Acme’s CFO. Bob confirms the relationship and agrees to an introduction. The observable change is from a broad question about the company’s network to a specific, approved introduction; the tool supplies the discovery step, while Bob supplies consent for the next action.

- **Relationship lookup:** Broad mail access stays inside a purpose-built tool; the agent receives a limited relationship result and contacts the relevant colleague.
- **Draft placement:** Colleagues can put draft emails in someone’s inbox, reducing the clicks needed when sales staff would otherwise ask that person to write an introduction.

Town uses this kind of tool for some recurring user activities, with an opt-out. More participants make more connections available, creating the proposed network effect. But privacy is still a product decision: people must understand and accept the information a tool reveals. Humans design each tool and explain its behavior, so the approach remains largely manual rather than expanding automatically with model capability.

### Move shareable knowledge out of private silos

[8:43](https://www.youtube.com/watch?v=REascnFlq_8&t=523s)

The third strategy creates a shared place where useful information accumulates. Private records can stay private while material suitable for the company or a subgroup moves into a wiki, database, or other common store. All relevant agents can search that store. The benefit comes from making already-shareable knowledge available beyond the conversation or inbox where it originated.

- **Shared skills:** An engineer saves a better database-profiling procedure in the repository. When another engineer encounters a slow query, their agent can use that procedure instead of rediscovering it.
- **Shared records:** Agents contribute information to a wiki or Airtable over time. This requires both tools for reading and writing the store and instructions or incentives that make those actions part of ordinary work.

A sweeper AI makes contribution a separate, recurring job. An agent inside each private silo knows what must remain there and knows the available shared destinations. At the end of the day, it examines new information and moves suitable contributions into those spaces. “Public” here means available within the company, rather than published on the internet. [10:13](https://www.youtube.com/watch?v=REascnFlq_8&t=613s)

Who decides which contributions may leave? Two versions put the decision in different hands:

- **Human approval:** The model prepares a list of proposed contributions and asks the user to approve publication to the shared space. It removes much of the writing work while retaining a review step.
- **Policy enforcement:** The model applies a policy itself and publishes the information it judges permissible. That removes routine approvals but makes disclosure quality depend on the model’s application of the policy.

What must the sweeper know to move information safely? The diagram separates the new private material from the rules and destination descriptions used to evaluate it. Both approval paths lead to the same shared store; they differ in who authorizes the contribution. This is where the shared-silo approach can begin to benefit from better judgment by the model.

Greze forecasts adoption of automatic policy enforcement within the next six months, especially in high-trust companies with ten or fifty employees and clear exclusions such as finance and HR data. He expects large enterprises to move more slowly. This is his adoption forecast, not evidence that the disclosure problem has already been solved.

How a sweeper contributes private knowledge
A sweeper evaluates new material using retention rules and shared-space descriptions. A human or the model’s policy judgment authorizes what leaves the silo.
- New private information — Material accumulated inside a private silo.
- Policy and destinations — What must stay private, and descriptions of the shared spaces.
- Daily sweeper — Examines new information and selects contributions.
- Human approval — The user reviews proposed contributions.
- Automatic policy decision — The LLM decides what the policy permits.
- Company shared spaces — Contributions become available to other agents.
- New private information → Daily sweeper (Examine new material)
- Policy and destinations → Daily sweeper (Guide selection)
- Daily sweeper → Human approval (Propose contributions)
- Daily sweeper → Automatic policy decision (Apply policy)
- Human approval → Company shared spaces (Publish approved material)
- Automatic policy decision → Company shared spaces (Publish permitted material)

### Search first, ask the person whose information matters

[11:43](https://www.youtube.com/watch?v=REascnFlq_8&t=703s)

The fourth strategy keeps humans in the information path. Your agent asks another person’s agent about an Acme finance connection. That person approves searching their email, sees the result, and separately approves sending it back. Consent precedes both discovery and disclosure.

This becomes expensive when only a few people hold the answer. In the hypothetical hundred-person company, the question can trigger a hundred Slack requests for permission. Most people must respond simply to establish that their information is irrelevant. Human attention becomes part of the search cost.

The fifth strategy, the black box, changes the timing. A trusted LLM can search across private silos without asking each owner first, while its trace remains inaccessible to ordinary participants. It stops at the answer-sharing point or before a write tool call, identifies the information needed for that output or action, and requests approval from the people who own it. Greze presents this as a powerful design he has rarely seen in practice.

Return to Bob and Jane. The request goes to all the company’s agents, which search Gmail and other private stores automatically. The black-box agent receives twenty matching connections, compares their strength using the email contacts, and selects Bob. Only then does Bob receive a prompt asking whether his connection to Jane may be shared with Jean-Denis for an introduction. The change from the previous approach is visible: discovery still spans a hundred people’s information, but the approval request reaches the selected owner.

Where does human attention enter this search? The diagram places Bob’s approval after candidate discovery and ranking, before disclosure. That saves irrelevant approval requests by giving the internal search much broader access. The privacy decision has moved later; the need to trust the search system has grown.

The final output must contain only information approved for sharing. Even a narrow relationship answer can expose something sensitive: asking about a connection to a recruiter at another company could reveal that a colleague is interviewing there. Hiding the search trace does not remove that inference. The approval step and the content allowed through it need careful design.

The Acme introduction through a black box
In the hypothetical example, broad automatic search finds twenty candidates, ranking selects Bob, and Bob approves sharing his connection before disclosure.
- Find an Acme CFO connection — Jean-Denis asks the company-wide question.
- Search private silos — Company agents inspect Gmail and private information without individual search approvals.
- Twenty connections — The black-box agent receives the matching candidates.
- Select Bob — Compare connection strength using the email contacts.
- Ask Bob — May his connection to Jane be shared for the introduction?
- Approved disclosure — Share the approved relationship information with Jean-Denis.
- Find an Acme CFO connection → Search private silos (Distribute request to agents)
- Search private silos → Twenty connections (Return matching connections internally)
- Twenty connections → Select Bob (Compare strength)
- Select Bob → Ask Bob (Request selected owner’s consent)
- Ask Bob → Approved disclosure (Proceed after approval)

### Shared memory can preserve the wrong thing

[15:14](https://www.youtube.com/watch?v=REascnFlq_8&t=914s)

Greze’s immediate-return bet is the automatically maintained information base, whether implemented as a wiki or database. Its value is straightforward: information becomes available for later work without someone repeatedly copying it between private conversations and shared records. But the same persistence that makes it useful can keep errors alive.

His personal wiki still calls his agent Apex, although he renamed it Ivy a month earlier. The old name survives somewhere in the memory setup and keeps resurfacing. A wrong agent name is funny; a wrong business fact carried forward into future decisions is harder to dismiss. The example illustrates stale information persisting, without identifying the particular memory mechanism responsible.

- **Prompt injection:** Material planted in a more open silo can enter context through agentic search and influence subsequent behavior.
- **Persistent factual errors:** A model mistake can enter the shared wiki and continue supplying wrong information to later work.
- **Wrong disclosures:** Automatic decisions can release information inappropriately, with consequences ranging from negligible harm to firing or a customer lawsuit.

The black box also creates an operational tension. Someone will eventually need to audit what happened. Greze asks who approves what, what gets logged, and what can be reversed. A system whose traces are hidden from requesters still needs an accountable way to inspect its data use, potentially through someone in the CISO organization. Secrecy from ordinary users does not eliminate the need for internal oversight.

### Start with a low-sensitivity zone

[17:15](https://www.youtube.com/watch?v=REascnFlq_8&t=1035s)

The proposed frontier is automatic judgment about when permission is necessary. Greze compares it to coding workflows that moved from approving every action, through “YOLO” execution, toward an auto mode that intervenes selectively. For information sharing, the corresponding design separates low-sensitivity material that may move automatically from material that needs human review or always requires approval.

Supplier meeting notes make the decision concrete. A user asks for notes from a weekly recurring call with a supplier and its finance team. In the proposed system, the model examines the notes, considers the requester’s role, and decides whether disclosure is acceptable without asking everyone involved. The presence of finance participants alone does not settle the decision; content and role inform the risk judgment.

This design gives model improvements somewhere useful to go. Better judgment, clearer policies, and carefully designed tools could expand the set of sharing decisions handled automatically, including decisions about when to involve a human. Greze’s recommendation is to define a low-sensitivity zone where the organization already accepts model discretion, then develop from there. Expansion remains conditional on those improvements; a more capable model does not itself grant permission to disclose more.

The desired outcome is less human coordination work as the system improves. Greze gives that ambition a pleasantly concrete destination: a beach in Hawaii with his kids. The architectural question underneath it is whether better models can take over more of the information-routing work, rather than leave humans maintaining the same silos and approving the same requests.

### The harder network effect crosses company lines

[19:15](https://www.youtube.com/watch?v=REascnFlq_8&t=1155s)

Within a company, shared governance offers a starting point for agreement about access and disclosure. The larger opportunity is getting several companies to accept a common agent working across their private information. That requires a use case valuable enough for each participant to permit some cooperation while retaining restrictions on what may be accessed or shared.

Greze describes an unnamed finance company exploring this direction with investment banks. The potential benefit is sharing private information about private companies for activities such as lending, with agents helping decide what can be accessed. The example is exploratory and unnamed, so it supports a possible beachhead rather than a proven deployment. Its appeal is that another institution’s participation could make useful information available that no single participant holds alone.

The ending leaves a real question open: how comfortable should people become with agents deciding privacy? Greze expects more decisions to move in that direction, yet does not claim to trust a future in which agents make all of them. The opportunity and the hesitation belong together. Work across silos can become more useful as judgment improves, while every new connection still needs an acceptable answer to who may learn what.

## Key takeaways

- Evaluate agent collaboration by whether it brings the right information into the consequential model call. Privacy limits that search even when context capacity is unlimited.
- Custom tools can search broadly while returning limited results, such as relationship scores. Their privacy tradeoffs still require deliberate design and user acceptance.
- A sweeper agent turns private work into proposed shared knowledge using retention policies and destination descriptions. Human approval and automatic policy enforcement offer different levels of disclosure control.
- Searching before requesting consent can spare irrelevant owners an approval prompt, but it requires trusting broad internal access and controlling the final disclosure.
- Begin automatic sharing with an accepted low-sensitivity zone. Persistent errors, prompt injection, disclosure mistakes, and audit needs remain part of the design as that zone grows.

## Related talks

- [Agentic Search for Context Engineering](https://ai.engineer/talks/ynJyIKwjonM-agentic-search-context-engineering) — Continues the opening question of how search supplies useful context before an agent acts.
- [Don't Build Agents, Build Skills Instead](https://ai.engineer/talks/CEvIs9y1uog-agent-skills) — Develops reusable procedural knowledge, one of the shared-space examples in this recording.
- [Your company brain will leak secrets: how we stopped it for big banks — Tanmai Gopal, PromptQL](https://ai.engineer/talks/0uC6u0lJJl4-your-company-brain-will-leak-secrets) — Offers a closely related topic for the disclosure risks of shared company knowledge and the closing investment-bank example.

## Slides

### Agents’ Next Frontier

Jean-Denis Greze introduces his view that agents working together can create network effects as they help people get things done.

[0:28](https://www.youtube.com/watch?v=REascnFlq_8&t=28s) [Title slide for a talk about agent-to-agent systems and network effects.](https://assets.ai.engineer/sites/aiecode2025/sha256/488547b06d63453606da53259a27a428ac5a6b608cda06b5cc66e8cc38ab9ba4.webp)

AI ENGINEER WORLD’S FAIR 2026

Agents’ Next Frontier
Agent-to-Agent & Network Effects

Jean-Denis Greze · Town · Tuesday, June 30, 2026 · https://x.com/jgreze · [redacted]

Interface text:
TOWN
View · Download · Share · Assistant Chat
AIE WF26 — Agents’ next fro...
1:30 PM - 1:50 PM
Take notes

### Running an agent is a search problem

Greze frames agent execution as a search for the right context before a consequential LLM call. The diagram traces one path through a branching search tree to the final call.

[1:11](https://www.youtube.com/watch?v=REascnFlq_8&t=71s) [Slide describes finding the right context as a search problem, illustrated by paths through a search tree.](https://assets.ai.engineer/sites/aiecode2025/sha256/8de60d9c47a698175cc1b7055ef494f75268f8ebd66ccde5cc3ebda27f5bbc2c.webp)

AGENT AS SEARCH

Running an agent is a search problem

- Running an AI system = a search problem: searching for the right context to put in front of the model
- LLM calls that produce the externally-visible outcomes are the ones that matter (a tool that changes state, or a result shown to the user)
- Everything before it just populates the context window: manual → RAG → agentic search
- Most known approaches are search-like ways to get the right info into the window

Diagram:
running an agent = search
START
LAST LLM CALL
FIG. 1 — running an agent = search

Interface text:
TOWN
View
Download
Share
Assistant Chat
AIE WF26 — Agents’ next fro...
1:30 PM - 1:50 PM
Take notes

### Multi-agent: one context window, many silos

Greze frames multi-agent systems as a way to bring information from separate people and organizations into one context window. His ideal of access to all data runs up against privacy and security.

[2:31](https://www.youtube.com/watch?v=REascnFlq_8&t=151s) [Diagram shows people and organizations supplying information to one context window, alongside a discussion of privacy and transaction costs.](https://assets.ai.engineer/sites/aiecode2025/sha256/e94d4116d8529f759c394361e5039890f2255448968c348a9c6a4ee5e05fe49e.webp)

THE IDEAL STATE

Multi-agent = one context window, many silos

– The intuitive framing — "many agents each doing things" — is misleading
– Better: a multi-agent system is a way to populate ONE context window from many sources / silos — across people, companies, governments with the information that leads to the desired outcome
– Ideal world: one agent with access to ALL the world's data gets the right outcome

Diagram labels: PERSON A; PERSON B; COMPANY X; CONTEXT WINDOW; COMPANY Y; ANOTHER ORG

FIG. 2 — one context window, many silos

Coase Theorem: with no transaction cost to bargaining, parties reach the efficient outcome on their own. Privacy & security are the transaction cost — that's the whole game.

Interface text: TOWN; View; Download; Share; Assistant Chat

### Five strategies to get the LLM the right data

Greze maps five ways to get siloed data into an LLM: bounded access, custom-tool privacy tradeoffs, shared silos, humans relaying information, and human approval of agent disclosures.

[4:11](https://www.youtube.com/watch?v=REascnFlq_8&t=251s) [A numbered list outlines five ways to give an LLM access to information across silos.](https://assets.ai.engineer/sites/aiecode2025/sha256/4604da24c7cbeea0f7760cd9d84ad5c9c3b70467cca97762acc478ef0fbf8367.webp)

Five strategies to get the LLM the right data

BREAKING THROUGH SILOS

1. Give the LLM access to all the info (within a trust boundary)
2. Build custom tools with a power-to-privacy tradeoff you're OK with
3. Create shared silos that accumulate info over time
4. Use humans as the conduit across silos
5. Let agents work in a black box, then HITL-approve the disclosure

Interface text:
TOWN
View
Download
Share
Assistant Chat

### Access within trust boundaries

Greze’s household agent accesses both spouses’ systems; an HR team could give an agent similar access. He argues this manual approach still relies on humans to organize data and does not improve as models get smarter.

[4:22](https://www.youtube.com/watch?v=REascnFlq_8&t=262s) [Slide describes shared agents with broad access within personal or workplace trust boundaries and notes limitations of manual access management.](https://assets.ai.engineer/sites/aiecode2025/sha256/665bd853792ac78e5e32b04b4b755b0a16dab2a7d6e43e894ac81b320cc94e00.webp)

STRATEGY 1

Approximate ”access to everything” within trust boundaries

– Personal: spouses' shared agent — email, calendar, Dropbox. Nothing to hide; runs the household better.
– Work: an HR team agent that can access everything the HR team can. Treat it like an employee; only HR interacts with it.
– This is the world IT / Security knows best — but it's manual and not dynamic (requires humans + de-silo-ification does not benefit from more intelligence)

If this is your only strategy, you lose hard over the next 2 years

Interface text:
TOWN
View
Download
Share
Assistant Chat

### Privacy-preserving cross-silo tools

A proposed tool could search employees’ Gmail for useful connections and return a relationship score without exposing conversation contents. Another would let colleagues put draft emails in your inbox.

[5:56](https://www.youtube.com/watch?v=REascnFlq_8&t=356s) [Slide describes tools that share limited signals across silos; an illustration shows obscured messages producing a relationship score.](https://assets.ai.engineer/sites/aiecode2025/sha256/8bccb052a96d6618d964992c6fe139c8d4a4012a653d00b25c858d5ff003507f.webp)

STRATEGY 2

Privacy-preserving cross-silo tools

– Build tools that share info across silos in a privacy-preserving way
– Use case: "Is anyone at my company connected to someone on the finance team at Acme Corp?"
– Relationship-strength tool: works across each employee's Gmail, return only a score (strength, recency, volume) — never conversation contents
– Another: a tool that lets anyone at your company put a draft in your inbox
– Pattern: preserve enough privacy to meet people's expectations
– This strategy and the previous one, trust boundaries, is what FDE armies at OpenAI / Google / Anthropic do — it maps cleanly to how SaaS works

Only this? You lose hard over the next 2 years

Diagram:
SIGNAL, NOT CONTENT
private messages
RELATIONSHIP SCORE: 0.92
recency
volume
hide the content, expose only the score.
FIG. 3 — signal, not content

Interface text:
TOWN
View
Download
Share
Assistant Chat

### Create shared silos

A proposed AI sweeper distills information from private silos into shared workspaces, where agents can benefit from collective knowledge. The open question is what private information to share and who decides.

[8:12](https://www.youtube.com/watch?v=REascnFlq_8&t=492s) [A slide describing shared knowledge silos, trusted AI sweepers, and decisions about what private information to share.](https://assets.ai.engineer/sites/aiecode2025/sha256/f52c4b4710776e0d8c5265cc3a7b9cb5d58c1e3f432bd6a5ae99b84167d96ab6.webp)

STRATEGY 3
Create shared silos

- Shared workspaces where data accumulates — within a company, or across companies
- Examples: Shared skills, encourage LLMs to accumulate knowledge in a shared mediums (your wiki, an Airtable, etc.)
- If everyone contributes and has tools and trajectories that emphasize getting data out, everyone's agents benefit and get better
- Best-in-class: a trusted "sweeper" AI inside each private silo that distills and pushes the right data up into the shared silo. This is LLM-wikis/brains but for teams.
- But how do we pick which private information is ok to share in a shared silo? Ask the user each time, or run an LLM-enforced policy?

Prediction: huge game-changer in the next 6 months — and it gets easier & better as models get smarter.

Interface text:
TOWN
View
Download
Share
Assistant Chat

### Privacy-preserving cross-silo tools

A proposed tool would score relationships across employees’ Gmail while hiding conversation contents. Greze says the approach remains manual and people may object to its privacy trade-off.

[8:18](https://www.youtube.com/watch?v=REascnFlq_8&t=498s) [A slide proposing cross-silo tools that expose relationship scores while hiding message contents.](https://assets.ai.engineer/sites/aiecode2025/sha256/54848989918e42f224464bf212f4fd16f985b6d5eb474358cc7a809273872130.webp)

STRATEGY 2
Privacy-preserving cross-silo tools

— Build tools that share info across silos in a privacy-preserving way
— Use case: "Is anyone at my company connected to someone on the finance team at Acme Corp?"
— Relationship-strength tool: works across each employee's Gmail, return only a score (strength, recency, volume) — never conversation contents
— Another: a tool that lets anyone at your company put a draft in your inbox
— Pattern: preserve enough privacy to meet people's expectations
— This strategy and the previous one, trust boundaries, is what FDE armies at OpenAI / Google / Anthropic do — it maps cleanly to how SaaS works

Only this? You lose hard over the next 2 years

Diagram:
SIGNAL, NOT CONTENT
private messages
RELATIONSHIP SCORE: 0.92
recency
volume
hide the content, expose only the score.
FIG. 3 – signal, not content

Interface text:
TOWN
View
Download
Share
Assistant Chat

### Create shared silos

A shared silo lets agents build on knowledge contributed across a team. A proposed sweeper AI would move useful information from private silos into shared spaces; the hard question is what private information to share.

[8:39](https://www.youtube.com/watch?v=REascnFlq_8&t=519s) [A slide describing shared knowledge silos, trusted AI sweepers, and decisions about what private information to share.](https://assets.ai.engineer/sites/aiecode2025/sha256/b480245e28996520377361adf402c62d4a9f44dfcaf97ae0ace60cd281a6cdb4.webp)

STRATEGY 3
Create shared silos

— Shared workspaces where data accumulates — within a company, or across companies
— Examples: Shared skills, encourage LLMs to accumulate knowledge in a shared mediums (your wiki, an Airtable, etc.)
— If everyone contributes and has tools and trajectories that emphasize getting data out, everyone's agents benefit and get better
— Best-in-class: a trusted "sweeper" AI inside each private silo that distills and pushes the right data up into the shared silo. This is LLM-wikis/brains but for teams.
— But how do we pick which private information is ok to share in a shared silo? Ask the user each time, or run an LLM-enforced policy?

Prediction: huge game-changer in the next 6 months — and it gets easier & better as models get smarter.

Interface text:
TOWN
View
Download
Share
Assistant Chat

### Humans as the conduit

Each data owner decides whether their agent can work on a request and share the result. The tradeoff: a request sent to 100 coworkers can interrupt everyone, even when few have useful information.

[11:45](https://www.youtube.com/watch?v=REascnFlq_8&t=705s) [A slide describing data owners as gatekeepers for agent requests and the risk of excessive approval requests.](https://assets.ai.engineer/sites/aiecode2025/sha256/eb4381c69a22e5d08a0824ba2b3d543fc00f96f5029d801612a93f24da6eb269.webp)

STRATEGY 4
Humans as the conduit

- Don't automate the flow — route it through the humans who own the data.
- Example: User A says "Can you ask all my co-workers' agents if are connected to someone in the finance team at ACME Corp?"
- The agent surfaces the request to each co-worker/s agent; each co-worker decides whether it's ok for their agent to work on the request. This is often what people think of as traditional A2A.
- The "no" is a feature, not a failure — the owner is the gatekeeper over their own agent (and all of the owners' data).
- Failure mode: the dumb version is spammy (ask 100 people up front). Nobody wants that.

“Draft an intro from my CEO to ABCCo's CEO (whom he met at AIE World's Fair), pitching our new product, cc me.”

Interface text:
TOWN
View
Download
Share
Assistant Chat

### Black-box compute and approval

Agents could search across private silos, then ask only the relevant data owners to approve disclosure. The approach depends on trusting the black box and getting the human approval step right.

[12:41](https://www.youtube.com/watch?v=REascnFlq_8&t=761s) [A slide proposing computation across silos followed by human approval before a worthwhile result is disclosed.](https://assets.ai.engineer/sites/aiecode2025/sha256/46775ef129f5168a4eea0e2d33b19d51dbaad1938e6f0e6b50077460e1bafbea.webp)

STRATEGY 5 — BETTER A2A

Black-box compute, and only ask to approve the inputs if the output is worthwhile

– Let agents read broadly from silos in a black box, but stop for HITL before any write / disclosure — and only once there's a result worth disclosing
– Example: User A says "Can you ask all my co-workers' agents if are connected to someone in the finance team at ACME Corp?"
– Instead of asking everyone up front, ask only the few with strong relationship strength, figure out who does and then only ask those 5 out of 100 people" "Can I reply to JDG letting them know that you're connected to the CFO at ACME Corp and willing to do an intro?"
– The insight here is that if you trust the black box, you can do all of the work across silos and only ask the people whose data was used to compute the result whether they are ok sharing back the result.
– What A2A looks like: Agent A → asks → Agent B → does work → if result is good, HITL approval → replies → Agent A

Interface text: TOWN; View; Download; Share; Assistant Chat

### The frontier is auto

Greze proposes starting with automatic sharing of low-sensitivity information, while retaining human review and approval for other cases. As models improve, he expects the automatic zone to expand.

[17:29](https://www.youtube.com/watch?v=REascnFlq_8&t=1049s) [Slide describing a shift toward automated decisions, with concentric zones for automatic handling, human review, and human approval.](https://assets.ai.engineer/sites/aiecode2025/sha256/8decdb86fcd4d73f8e2680af268e509baea44f050eaaa1e589b3d3b5dc6afb4b.webp)

THE FRONTIER
The frontier is auto

Coding  Approvals → Skip → Auto
A2A across silos  HITL → Auto

– Auto for (i) building shared silos and (ii) eventually auto-deciding when to even ask the human
– Scales with model capability — the smarter the model, the better it models human privacy expectations & org policy

“Notes from a weekly recurring call with someone on a supplier's finance team are fine to share — I won't even HITL you.”

Diagram:
Auto
Low sensitivity
Human review
Always human-approved
automation grows from the edges in
FIG. 4 — automation expands outward-in

Interface text:
TOWN
View
Download
Share
Assistant Chat

### Network effects

Greze frames A2A approaches as ways to bring information across silos into one context window. He sees privacy expectations and human review as limits, and asks whether agents can make privacy decisions across companies.

[19:10](https://www.youtube.com/watch?v=REascnFlq_8&t=1150s) [Closing slide summarizing cross-silo information sharing, agent capabilities, privacy limits, and questions about automated privacy decisions.](https://assets.ai.engineer/sites/aiecode2025/sha256/767e36a304cffabbbd4f83a20f168030b3cd6c253e32f284cf8f3fa5289a7163.webp)

CLOSING

Network effects

– Trust boundaries · custom tools · shared silos · humans · black-box — all just strategies to get the right info into one context window across silos
– From our own usage, A2A makes systems extremely powerful, but it's still difficult for humans to wrap their head around what's going. Shared silos is the most straight forward model for people to understand. Black-box is super powerful, but it's tough to grasp.
– The main limiter is human expectations around privacy. And every HITL slows things down.
– As the economy lets humans do more with AI, agents get access to more data — including more public-ish datasets without these limits. Ultimately even privacy will be moderated by auto mode that's aware of individuals' and organizations' policies.
– Open question: all of these network effects within companies are happening now, what does it look like across companies and what are the first use cases for "auto-privacy/security" across companies?

Can we trust a future where we trust agents to make decisions around privacy?

Interface text:
TOWN
View
Download
Share
Assistant Chat

### AI Engineer World's Fair

[21:10](https://www.youtube.com/watch?v=REascnFlq_8&t=1270s) [AI Engineer World's Fair title in white on a black background.](https://assets.ai.engineer/sites/aiecode2025/sha256/c2a3026f032138390584bcb5cb198b8ca17e2fb51d037515237f8a7d75dd8cdd.webp)

AI Engineer
World’s Fair

HTTPS://AI.ENGINEER

## Full transcript

[0:01](https://www.youtube.com/watch?v=REascnFlq_8&t=1s) [music]

[0:12](https://www.youtube.com/watch?v=REascnFlq_8&t=12s) >> Can you all hear me? All right.

[0:14](https://www.youtube.com/watch?v=REascnFlq_8&t=14s) Um

[0:15](https://www.youtube.com/watch?v=REascnFlq_8&t=15s) Well, first things for coming. I can't

[0:17](https://www.youtube.com/watch?v=REascnFlq_8&t=17s) believe there's anybody in the room, but

[0:18](https://www.youtube.com/watch?v=REascnFlq_8&t=18s) that's very nice. Uh

[0:20](https://www.youtube.com/watch?v=REascnFlq_8&t=20s) Uh my name is Jean-Denis. Um I'm CTO at

[0:23](https://www.youtube.com/watch?v=REascnFlq_8&t=23s) a company called Town. We're not going

[0:25](https://www.youtube.com/watch?v=REascnFlq_8&t=25s) to really talk about Town, so you can go

[0:27](https://www.youtube.com/watch?v=REascnFlq_8&t=27s) to town.com and check that out if you

[0:29](https://www.youtube.com/watch?v=REascnFlq_8&t=29s) want, but that's not the point of the

[0:30](https://www.youtube.com/watch?v=REascnFlq_8&t=30s) talk today. I was CTO at Plaid for 7

[0:33](https://www.youtube.com/watch?v=REascnFlq_8&t=33s) years, and then I was at Dropbox before.

[0:36](https://www.youtube.com/watch?v=REascnFlq_8&t=36s) And then before that I built software

[0:38](https://www.youtube.com/watch?v=REascnFlq_8&t=38s) for hedge funds. I've done lots of stuff

[0:39](https://www.youtube.com/watch?v=REascnFlq_8&t=39s) in my career, and right now I'm working

[0:42](https://www.youtube.com/watch?v=REascnFlq_8&t=42s) on

[0:43](https://www.youtube.com/watch?v=REascnFlq_8&t=43s) uh assistance agents for for normal

[0:45](https://www.youtube.com/watch?v=REascnFlq_8&t=45s) people, not not for engineers, but for

[0:47](https://www.youtube.com/watch?v=REascnFlq_8&t=47s) like basically everyone in America and

[0:48](https://www.youtube.com/watch?v=REascnFlq_8&t=48s) the world.

[0:50](https://www.youtube.com/watch?v=REascnFlq_8&t=50s) And one of the things we've been working

[0:51](https://www.youtube.com/watch?v=REascnFlq_8&t=51s) on are systems where agents work with

[0:54](https://www.youtube.com/watch?v=REascnFlq_8&t=54s) other agents. So, agent-to-agent. And

[0:56](https://www.youtube.com/watch?v=REascnFlq_8&t=56s) the main idea is that we think there's

[0:59](https://www.youtube.com/watch?v=REascnFlq_8&t=59s) huge network effects if agents can work

[1:02](https://www.youtube.com/watch?v=REascnFlq_8&t=62s) together to get things done for people

[1:04](https://www.youtube.com/watch?v=REascnFlq_8&t=64s) because in the real world the way most

[1:06](https://www.youtube.com/watch?v=REascnFlq_8&t=66s) of us do work is with other people,

[1:09](https://www.youtube.com/watch?v=REascnFlq_8&t=69s) right? Uh

[1:10](https://www.youtube.com/watch?v=REascnFlq_8&t=70s) more is better. But, actually

[1:13](https://www.youtube.com/watch?v=REascnFlq_8&t=73s) I don't think agent-to-agent makes much

[1:15](https://www.youtube.com/watch?v=REascnFlq_8&t=75s) sense as a concept. So, I want to

[1:17](https://www.youtube.com/watch?v=REascnFlq_8&t=77s) reframe the entire talk in terms of

[1:19](https://www.youtube.com/watch?v=REascnFlq_8&t=79s) search. So, I think

[1:22](https://www.youtube.com/watch?v=REascnFlq_8&t=82s) most LLM systems are just a search

[1:24](https://www.youtube.com/watch?v=REascnFlq_8&t=84s) problem.

[1:25](https://www.youtube.com/watch?v=REascnFlq_8&t=85s) And what you're trying to do is you're

[1:27](https://www.youtube.com/watch?v=REascnFlq_8&t=87s) trying to make sure the context window

[1:30](https://www.youtube.com/watch?v=REascnFlq_8&t=90s) right before you either return results

[1:32](https://www.youtube.com/watch?v=REascnFlq_8&t=92s) to the user or before a tool call,

[1:34](https://www.youtube.com/watch?v=REascnFlq_8&t=94s) you're trying to make sure the context

[1:35](https://www.youtube.com/watch?v=REascnFlq_8&t=95s) window has the right information for the

[1:36](https://www.youtube.com/watch?v=REascnFlq_8&t=96s) user. If you put the right information

[1:38](https://www.youtube.com/watch?v=REascnFlq_8&t=98s) in the context window

[1:40](https://www.youtube.com/watch?v=REascnFlq_8&t=100s) then based on the

[1:42](https://www.youtube.com/watch?v=REascnFlq_8&t=102s) intelligence, so to speak, of the LLM,

[1:44](https://www.youtube.com/watch?v=REascnFlq_8&t=104s) you will get the best result possible.

[1:46](https://www.youtube.com/watch?v=REascnFlq_8&t=106s) Um so, you know, 4 years ago the way we

[1:48](https://www.youtube.com/watch?v=REascnFlq_8&t=108s) did that is humans would populate the

[1:51](https://www.youtube.com/watch?v=REascnFlq_8&t=111s) context window manually. Then a couple

[1:53](https://www.youtube.com/watch?v=REascnFlq_8&t=113s) years ago, most people are ragging, so

[1:55](https://www.youtube.com/watch?v=REascnFlq_8&t=115s) they were like, let's have a tool, like

[1:56](https://www.youtube.com/watch?v=REascnFlq_8&t=116s) a search tool, that can look across

[1:58](https://www.youtube.com/watch?v=REascnFlq_8&t=118s) systems and bring the data in there. And

[2:00](https://www.youtube.com/watch?v=REascnFlq_8&t=120s) then people were like, well, that

[2:01](https://www.youtube.com/watch?v=REascnFlq_8&t=121s) doesn't scale super well, has issues.

[2:04](https://www.youtube.com/watch?v=REascnFlq_8&t=124s) And now we're all about agentic search,

[2:05](https://www.youtube.com/watch?v=REascnFlq_8&t=125s) which is the idea that you give the

[2:07](https://www.youtube.com/watch?v=REascnFlq_8&t=127s) agent a lot of tools, and it'll search

[2:09](https://www.youtube.com/watch?v=REascnFlq_8&t=129s) through the space of all content,

[2:12](https://www.youtube.com/watch?v=REascnFlq_8&t=132s) and then hopefully before it makes a

[2:13](https://www.youtube.com/watch?v=REascnFlq_8&t=133s) tool call, it has exactly the right

[2:14](https://www.youtube.com/watch?v=REascnFlq_8&t=134s) content to make the right tool call to

[2:16](https://www.youtube.com/watch?v=REascnFlq_8&t=136s) return the right information to the

[2:18](https://www.youtube.com/watch?v=REascnFlq_8&t=138s) user.

[2:19](https://www.youtube.com/watch?v=REascnFlq_8&t=139s) Um and in this, by the way, there's no

[2:21](https://www.youtube.com/watch?v=REascnFlq_8&t=141s) there's no people. It's just a one LLM

[2:22](https://www.youtube.com/watch?v=REascnFlq_8&t=142s) call at the like the one that matters

[2:24](https://www.youtube.com/watch?v=REascnFlq_8&t=144s) having the right context. That's That's

[2:26](https://www.youtube.com/watch?v=REascnFlq_8&t=146s) what you're trying to do. You're trying

[2:27](https://www.youtube.com/watch?v=REascnFlq_8&t=147s) to engineer

[2:28](https://www.youtube.com/watch?v=REascnFlq_8&t=148s) um that system.

[2:30](https://www.youtube.com/watch?v=REascnFlq_8&t=150s) Cool. So, what what does that have to do

[2:31](https://www.youtube.com/watch?v=REascnFlq_8&t=151s) with agent agent? So, I want you to

[2:33](https://www.youtube.com/watch?v=REascnFlq_8&t=153s) imagine the following world.

[2:35](https://www.youtube.com/watch?v=REascnFlq_8&t=155s) There's

[2:37](https://www.youtube.com/watch?v=REascnFlq_8&t=157s) not many agents that can do things.

[2:39](https://www.youtube.com/watch?v=REascnFlq_8&t=159s) There's just one agent,

[2:41](https://www.youtube.com/watch?v=REascnFlq_8&t=161s) right? And it has one context window,

[2:43](https://www.youtube.com/watch?v=REascnFlq_8&t=163s) and it has access to all the information

[2:45](https://www.youtube.com/watch?v=REascnFlq_8&t=165s) in the universe.

[2:47](https://www.youtube.com/watch?v=REascnFlq_8&t=167s) It can look at any one person's email,

[2:49](https://www.youtube.com/watch?v=REascnFlq_8&t=169s) can look at any company's information,

[2:52](https://www.youtube.com/watch?v=REascnFlq_8&t=172s) can look at any government's

[2:53](https://www.youtube.com/watch?v=REascnFlq_8&t=173s) information,

[2:54](https://www.youtube.com/watch?v=REascnFlq_8&t=174s) and it has it right there in the context

[2:55](https://www.youtube.com/watch?v=REascnFlq_8&t=175s) window,

[2:56](https://www.youtube.com/watch?v=REascnFlq_8&t=176s) and then you ask it to do something. You

[2:59](https://www.youtube.com/watch?v=REascnFlq_8&t=179s) You have your little system prompt with

[3:00](https://www.youtube.com/watch?v=REascnFlq_8&t=180s) all that data, and what's going to

[3:02](https://www.youtube.com/watch?v=REascnFlq_8&t=182s) happen is it'll give you the best

[3:03](https://www.youtube.com/watch?v=REascnFlq_8&t=183s) possible outcome.

[3:05](https://www.youtube.com/watch?v=REascnFlq_8&t=185s) And that actually That is a multi-agent

[3:08](https://www.youtube.com/watch?v=REascnFlq_8&t=188s) world.

[3:09](https://www.youtube.com/watch?v=REascnFlq_8&t=189s) It's just an agent that has access to

[3:10](https://www.youtube.com/watch?v=REascnFlq_8&t=190s) all the world's information. That's the

[3:12](https://www.youtube.com/watch?v=REascnFlq_8&t=192s) natural state of things. That's the

[3:13](https://www.youtube.com/watch?v=REascnFlq_8&t=193s) ideal state of things. There's a problem

[3:15](https://www.youtube.com/watch?v=REascnFlq_8&t=195s) with this state of things, and the

[3:17](https://www.youtube.com/watch?v=REascnFlq_8&t=197s) problem comes from

[3:19](https://www.youtube.com/watch?v=REascnFlq_8&t=199s) a few So, you're learning economics

[3:21](https://www.youtube.com/watch?v=REascnFlq_8&t=201s) something called the Coase theorem, and

[3:22](https://www.youtube.com/watch?v=REascnFlq_8&t=202s) it says that basically even humans, if

[3:24](https://www.youtube.com/watch?v=REascnFlq_8&t=204s) they all have access to all the right

[3:26](https://www.youtube.com/watch?v=REascnFlq_8&t=206s) information and there's no transaction

[3:27](https://www.youtube.com/watch?v=REascnFlq_8&t=207s) costs, we get the economically

[3:30](https://www.youtube.com/watch?v=REascnFlq_8&t=210s) ideal outcome out of a out of a contract

[3:32](https://www.youtube.com/watch?v=REascnFlq_8&t=212s) or negotiation. Well, it's the same

[3:34](https://www.youtube.com/watch?v=REascnFlq_8&t=214s) thing. We can't put all of the world's

[3:36](https://www.youtube.com/watch?v=REascnFlq_8&t=216s) contacts We can't make it available to

[3:38](https://www.youtube.com/watch?v=REascnFlq_8&t=218s) the LLM. Like theoretically even with

[3:40](https://www.youtube.com/watch?v=REascnFlq_8&t=220s) infinite context window, because of

[3:42](https://www.youtube.com/watch?v=REascnFlq_8&t=222s) privacy and security.

[3:44](https://www.youtube.com/watch?v=REascnFlq_8&t=224s) We're humans. I don't let you look at my

[3:46](https://www.youtube.com/watch?v=REascnFlq_8&t=226s) email, so there cannot be an agent that

[3:49](https://www.youtube.com/watch?v=REascnFlq_8&t=229s) I'm willing to just let it look at my

[3:50](https://www.youtube.com/watch?v=REascnFlq_8&t=230s) email all the time. But, if it existed,

[3:54](https://www.youtube.com/watch?v=REascnFlq_8&t=234s) it would be very, very powerful. So, I I

[3:56](https://www.youtube.com/watch?v=REascnFlq_8&t=236s) think this is like this is the test for

[3:58](https://www.youtube.com/watch?v=REascnFlq_8&t=238s) a multi-agent system, which is how well

[4:00](https://www.youtube.com/watch?v=REascnFlq_8&t=240s) does it approximate this?

[4:02](https://www.youtube.com/watch?v=REascnFlq_8&t=242s) If it approximates this, that means if

[4:03](https://www.youtube.com/watch?v=REascnFlq_8&t=243s) you can get the same data in your window

[4:06](https://www.youtube.com/watch?v=REascnFlq_8&t=246s) that a perfect system that has access to

[4:07](https://www.youtube.com/watch?v=REascnFlq_8&t=247s) all the world's data could, then you get

[4:09](https://www.youtube.com/watch?v=REascnFlq_8&t=249s) the optimal outcome. That's what you

[4:10](https://www.youtube.com/watch?v=REascnFlq_8&t=250s) need to try to do. So, we're going to

[4:12](https://www.youtube.com/watch?v=REascnFlq_8&t=252s) talk about five strategies that people

[4:14](https://www.youtube.com/watch?v=REascnFlq_8&t=254s) use at various companies to try to get

[4:17](https://www.youtube.com/watch?v=REascnFlq_8&t=257s) the right data into that that LLM call

[4:20](https://www.youtube.com/watch?v=REascnFlq_8&t=260s) with an externality. So,

[4:22](https://www.youtube.com/watch?v=REascnFlq_8&t=262s) the first one is

[4:25](https://www.youtube.com/watch?v=REascnFlq_8&t=265s) approximate access to everything within

[4:26](https://www.youtube.com/watch?v=REascnFlq_8&t=266s) a trust boundary.

[4:28](https://www.youtube.com/watch?v=REascnFlq_8&t=268s) So, my wife and I, we have an agent

[4:30](https://www.youtube.com/watch?v=REascnFlq_8&t=270s) together,

[4:31](https://www.youtube.com/watch?v=REascnFlq_8&t=271s) and that agent has access to my email

[4:33](https://www.youtube.com/watch?v=REascnFlq_8&t=273s) and her email,

[4:35](https://www.youtube.com/watch?v=REascnFlq_8&t=275s) uh including emails before we were

[4:37](https://www.youtube.com/watch?v=REascnFlq_8&t=277s) married.

[4:39](https://www.youtube.com/watch?v=REascnFlq_8&t=279s) Uh and it's okay, she doesn't ask my

[4:40](https://www.youtube.com/watch?v=REascnFlq_8&t=280s) agent questions about that, but she does

[4:41](https://www.youtube.com/watch?v=REascnFlq_8&t=281s) ask about, you know, whether I like

[4:43](https://www.youtube.com/watch?v=REascnFlq_8&t=283s) schedule something for our kids or, you

[4:45](https://www.youtube.com/watch?v=REascnFlq_8&t=285s) know, uh if I followed up on some

[4:47](https://www.youtube.com/watch?v=REascnFlq_8&t=287s) third-party thing. And so, the fact that

[4:49](https://www.youtube.com/watch?v=REascnFlq_8&t=289s) our agent has access to both of our

[4:50](https://www.youtube.com/watch?v=REascnFlq_8&t=290s) systems is wonderful.

[4:52](https://www.youtube.com/watch?v=REascnFlq_8&t=292s) Um and the work context, this might be

[4:54](https://www.youtube.com/watch?v=REascnFlq_8&t=294s) there's a an HR team agent that has

[4:56](https://www.youtube.com/watch?v=REascnFlq_8&t=296s) access to all the HR systems, just like

[4:58](https://www.youtube.com/watch?v=REascnFlq_8&t=298s) an employee of the HR team would, or

[5:00](https://www.youtube.com/watch?v=REascnFlq_8&t=300s) maybe as much access as the lowest

[5:02](https://www.youtube.com/watch?v=REascnFlq_8&t=302s) employee in the HR team. All the

[5:04](https://www.youtube.com/watch?v=REascnFlq_8&t=304s) employees in the HR team have the

[5:05](https://www.youtube.com/watch?v=REascnFlq_8&t=305s) ability to ask this agent questions, and

[5:08](https://www.youtube.com/watch?v=REascnFlq_8&t=308s) boom, it gets pretty good results.

[5:10](https://www.youtube.com/watch?v=REascnFlq_8&t=310s) And this is very popular right now.

[5:12](https://www.youtube.com/watch?v=REascnFlq_8&t=312s) Uh it's very popular with IT teams and

[5:14](https://www.youtube.com/watch?v=REascnFlq_8&t=314s) security teams cuz it's the same model

[5:17](https://www.youtube.com/watch?v=REascnFlq_8&t=317s) as SaaS for security, so it works really

[5:19](https://www.youtube.com/watch?v=REascnFlq_8&t=319s) well.

[5:20](https://www.youtube.com/watch?v=REascnFlq_8&t=320s) I think it has a problem, which is a

[5:22](https://www.youtube.com/watch?v=REascnFlq_8&t=322s) fundamental problem that if I wake up in

[5:24](https://www.youtube.com/watch?v=REascnFlq_8&t=324s) the morning, it's like basically the

[5:25](https://www.youtube.com/watch?v=REascnFlq_8&t=325s) only thing I think about, which is does

[5:27](https://www.youtube.com/watch?v=REascnFlq_8&t=327s) it get Over time, does this system

[5:30](https://www.youtube.com/watch?v=REascnFlq_8&t=330s) naturally require fewer humans?

[5:32](https://www.youtube.com/watch?v=REascnFlq_8&t=332s) And then, as the models get better, does

[5:35](https://www.youtube.com/watch?v=REascnFlq_8&t=335s) this approach get better? And the

[5:36](https://www.youtube.com/watch?v=REascnFlq_8&t=336s) problem with this approach is the answer

[5:37](https://www.youtube.com/watch?v=REascnFlq_8&t=337s) is no to both.

[5:39](https://www.youtube.com/watch?v=REascnFlq_8&t=339s) Uh you still need humans to think about

[5:41](https://www.youtube.com/watch?v=REascnFlq_8&t=341s) all the data, and you don't get magical

[5:43](https://www.youtube.com/watch?v=REascnFlq_8&t=343s) de-siloification of your data. You've

[5:45](https://www.youtube.com/watch?v=REascnFlq_8&t=345s) just created a new silo cuz a human

[5:47](https://www.youtube.com/watch?v=REascnFlq_8&t=347s) thought about it.

[5:49](https://www.youtube.com/watch?v=REascnFlq_8&t=349s) So, the problem with this is I do think

[5:50](https://www.youtube.com/watch?v=REascnFlq_8&t=350s) if this is your approach to building

[5:52](https://www.youtube.com/watch?v=REascnFlq_8&t=352s) better AI, you're going to be in

[5:55](https://www.youtube.com/watch?v=REascnFlq_8&t=355s) the next couple years.

[5:56](https://www.youtube.com/watch?v=REascnFlq_8&t=356s) Um but that's okay. Your is my

[5:59](https://www.youtube.com/watch?v=REascnFlq_8&t=359s) opportunity.

[6:00](https://www.youtube.com/watch?v=REascnFlq_8&t=360s) Uh I'm just I'm just not an I'm

[6:03](https://www.youtube.com/watch?v=REascnFlq_8&t=363s) sorry. That was mean. But like I think

[6:04](https://www.youtube.com/watch?v=REascnFlq_8&t=364s) it's not I think it's a good now way to

[6:06](https://www.youtube.com/watch?v=REascnFlq_8&t=366s) think about it. It's not the good end

[6:07](https://www.youtube.com/watch?v=REascnFlq_8&t=367s) game way to think about it.

[6:09](https://www.youtube.com/watch?v=REascnFlq_8&t=369s) The other approach which is I think is a

[6:10](https://www.youtube.com/watch?v=REascnFlq_8&t=370s) little more clever and then I'm going to

[6:11](https://www.youtube.com/watch?v=REascnFlq_8&t=371s) try to explain it is basically you try

[6:13](https://www.youtube.com/watch?v=REascnFlq_8&t=373s) to have tools that make a different

[6:15](https://www.youtube.com/watch?v=REascnFlq_8&t=375s) trade-off between power and privacy.

[6:18](https://www.youtube.com/watch?v=REascnFlq_8&t=378s) So, I'm going to give you an example

[6:19](https://www.youtube.com/watch?v=REascnFlq_8&t=379s) here. Um the use case is I want to ask

[6:22](https://www.youtube.com/watch?v=REascnFlq_8&t=382s) my agent, does anyone in my company is

[6:24](https://www.youtube.com/watch?v=REascnFlq_8&t=384s) anyone in my company connected to

[6:26](https://www.youtube.com/watch?v=REascnFlq_8&t=386s) someone on the finance team at Acme

[6:27](https://www.youtube.com/watch?v=REascnFlq_8&t=387s) Corp?

[6:29](https://www.youtube.com/watch?v=REascnFlq_8&t=389s) And so,

[6:31](https://www.youtube.com/watch?v=REascnFlq_8&t=391s) what the the no silo way to do that is

[6:33](https://www.youtube.com/watch?v=REascnFlq_8&t=393s) just give me access to everyone's Gmail

[6:34](https://www.youtube.com/watch?v=REascnFlq_8&t=394s) in my company.

[6:36](https://www.youtube.com/watch?v=REascnFlq_8&t=396s) I'll see who has emails with people from

[6:37](https://www.youtube.com/watch?v=REascnFlq_8&t=397s) Acme Corp. Then I'll look at their

[6:39](https://www.youtube.com/watch?v=REascnFlq_8&t=399s) profile on Google or LinkedIn and then

[6:42](https://www.youtube.com/watch?v=REascnFlq_8&t=402s) I'll be like, oh, you seem to email a

[6:43](https://www.youtube.com/watch?v=REascnFlq_8&t=403s) lot with the CFO. Can you do the intro

[6:45](https://www.youtube.com/watch?v=REascnFlq_8&t=405s) for me? But right, obviously silos, we

[6:47](https://www.youtube.com/watch?v=REascnFlq_8&t=407s) don't want that. So, what if you build a

[6:49](https://www.youtube.com/watch?v=REascnFlq_8&t=409s) tool and what the tool did is

[6:52](https://www.youtube.com/watch?v=REascnFlq_8&t=412s) it looked at everyone's Gmail. So, that

[6:53](https://www.youtube.com/watch?v=REascnFlq_8&t=413s) tool had access to everyone's Gmail and

[6:55](https://www.youtube.com/watch?v=REascnFlq_8&t=415s) it just returned a relationship strength

[6:57](https://www.youtube.com/watch?v=REascnFlq_8&t=417s) score.

[6:58](https://www.youtube.com/watch?v=REascnFlq_8&t=418s) So, the tool you would give it like a

[7:00](https://www.youtube.com/watch?v=REascnFlq_8&t=420s) domain and you would say I'm looking for

[7:02](https://www.youtube.com/watch?v=REascnFlq_8&t=422s) someone who's a CFO. It would look at

[7:04](https://www.youtube.com/watch?v=REascnFlq_8&t=424s) everyone at the company who sent emails

[7:06](https://www.youtube.com/watch?v=REascnFlq_8&t=426s) to that company and then it would like

[7:08](https://www.youtube.com/watch?v=REascnFlq_8&t=428s) rank their score and they would give you

[7:09](https://www.youtube.com/watch?v=REascnFlq_8&t=429s) back the score and then the agent would

[7:11](https://www.youtube.com/watch?v=REascnFlq_8&t=431s) get the score and it would be like cool.

[7:13](https://www.youtube.com/watch?v=REascnFlq_8&t=433s) Then they would use a Slack tool to text

[7:15](https://www.youtube.com/watch?v=REascnFlq_8&t=435s) that person the company. It's like, hey

[7:16](https://www.youtube.com/watch?v=REascnFlq_8&t=436s) Bob, I see that you're connected to the

[7:17](https://www.youtube.com/watch?v=REascnFlq_8&t=437s) Jane who's the CFO at Acme Corp. And

[7:20](https://www.youtube.com/watch?v=REascnFlq_8&t=440s) then Bob would be like, yes, I am. And

[7:22](https://www.youtube.com/watch?v=REascnFlq_8&t=442s) then your AI would be like, oh, can you

[7:23](https://www.youtube.com/watch?v=REascnFlq_8&t=443s) draft an e- can I draft an email or can

[7:25](https://www.youtube.com/watch?v=REascnFlq_8&t=445s) you draft an email introducing me? And

[7:26](https://www.youtube.com/watch?v=REascnFlq_8&t=446s) then Bob would say, yes, and he would do

[7:28](https://www.youtube.com/watch?v=REascnFlq_8&t=448s) that and you'd be connected and

[7:29](https://www.youtube.com/watch?v=REascnFlq_8&t=449s) everything would be wonderful. So, this

[7:31](https://www.youtube.com/watch?v=REascnFlq_8&t=451s) is actually very cool approach. I don't

[7:33](https://www.youtube.com/watch?v=REascnFlq_8&t=453s) know how many of you do it. Like we we

[7:35](https://www.youtube.com/watch?v=REascnFlq_8&t=455s) do this at Town for a few things that we

[7:37](https://www.youtube.com/watch?v=REascnFlq_8&t=457s) see a lot of our users do. We ask

[7:38](https://www.youtube.com/watch?v=REascnFlq_8&t=458s) ourself, what is a privacy preserving

[7:40](https://www.youtube.com/watch?v=REascnFlq_8&t=460s) tool that all of our users would be okay

[7:42](https://www.youtube.com/watch?v=REascnFlq_8&t=462s) existing? They can opt out if they don't

[7:43](https://www.youtube.com/watch?v=REascnFlq_8&t=463s) want it, but it has a natural network

[7:46](https://www.youtube.com/watch?v=REascnFlq_8&t=466s) effect because it breaks through silos

[7:47](https://www.youtube.com/watch?v=REascnFlq_8&t=467s) in interesting way.

[7:49](https://www.youtube.com/watch?v=REascnFlq_8&t=469s) Um, like another one that's interesting

[7:50](https://www.youtube.com/watch?v=REascnFlq_8&t=470s) here is letting other people put draft

[7:52](https://www.youtube.com/watch?v=REascnFlq_8&t=472s) emails in your inbox.

[7:54](https://www.youtube.com/watch?v=REascnFlq_8&t=474s) You let other people at your company

[7:55](https://www.youtube.com/watch?v=REascnFlq_8&t=475s) draft emails on your behalf cuz they're

[7:56](https://www.youtube.com/watch?v=REascnFlq_8&t=476s) going to ask you to anyway to get intros

[7:58](https://www.youtube.com/watch?v=REascnFlq_8&t=478s) if they're on the sales team, so might

[7:59](https://www.youtube.com/watch?v=REascnFlq_8&t=479s) as well save yourself a few clicks.

[8:02](https://www.youtube.com/watch?v=REascnFlq_8&t=482s) So, the question here is like are people

[8:03](https://www.youtube.com/watch?v=REascnFlq_8&t=483s) going to be okay with a privacy

[8:04](https://www.youtube.com/watch?v=REascnFlq_8&t=484s) trade-off that you make within a

[8:06](https://www.youtube.com/watch?v=REascnFlq_8&t=486s) corporation?

[8:08](https://www.youtube.com/watch?v=REascnFlq_8&t=488s) Uh,

[8:08](https://www.youtube.com/watch?v=REascnFlq_8&t=488s) bad. Bad.

[8:11](https://www.youtube.com/watch?v=REascnFlq_8&t=491s) Oh, boy.

[8:13](https://www.youtube.com/watch?v=REascnFlq_8&t=493s) Within a corporation that will, you

[8:14](https://www.youtube.com/watch?v=REascnFlq_8&t=494s) know, mostly it works. Um,

[8:17](https://www.youtube.com/watch?v=REascnFlq_8&t=497s) so the problem here again is it's again

[8:19](https://www.youtube.com/watch?v=REascnFlq_8&t=499s) manual and not dynamic. It's manual cuz

[8:21](https://www.youtube.com/watch?v=REascnFlq_8&t=501s) humans need to think about the tools.

[8:23](https://www.youtube.com/watch?v=REascnFlq_8&t=503s) Maybe I could build the tools.

[8:25](https://www.youtube.com/watch?v=REascnFlq_8&t=505s) Uh, and it's also manual cuz you need to

[8:27](https://www.youtube.com/watch?v=REascnFlq_8&t=507s) explain it to everyone that it's

[8:29](https://www.youtube.com/watch?v=REascnFlq_8&t=509s) happening. Humans may not like it if

[8:30](https://www.youtube.com/watch?v=REascnFlq_8&t=510s) this is happening if they're not okay

[8:32](https://www.youtube.com/watch?v=REascnFlq_8&t=512s) with the privacy security the privacy

[8:34](https://www.youtube.com/watch?v=REascnFlq_8&t=514s) kind of power trade-off that you've

[8:35](https://www.youtube.com/watch?v=REascnFlq_8&t=515s) made. Cool. And again, this doesn't

[8:38](https://www.youtube.com/watch?v=REascnFlq_8&t=518s) really get better as they I guess

[8:39](https://www.youtube.com/watch?v=REascnFlq_8&t=519s) better. That's the problem.

[8:41](https://www.youtube.com/watch?v=REascnFlq_8&t=521s) Cool. So, now the third category. This

[8:42](https://www.youtube.com/watch?v=REascnFlq_8&t=522s) one's super popular, but only mostly in

[8:45](https://www.youtube.com/watch?v=REascnFlq_8&t=525s) the single user context. So, this is,

[8:46](https://www.youtube.com/watch?v=REascnFlq_8&t=526s) you know, like personal wikis in claw

[8:48](https://www.youtube.com/watch?v=REascnFlq_8&t=528s) land. That's what we would call it. But

[8:50](https://www.youtube.com/watch?v=REascnFlq_8&t=530s) it's across teams. So,

[8:52](https://www.youtube.com/watch?v=REascnFlq_8&t=532s) it's a shared silo. Create a new place

[8:55](https://www.youtube.com/watch?v=REascnFlq_8&t=535s) where data accumulates

[8:56](https://www.youtube.com/watch?v=REascnFlq_8&t=536s) within your company within subgroups at

[8:58](https://www.youtube.com/watch?v=REascnFlq_8&t=538s) your company.

[8:59](https://www.youtube.com/watch?v=REascnFlq_8&t=539s) Um, and you start to put more and more

[9:01](https://www.youtube.com/watch?v=REascnFlq_8&t=541s) stuff there over time.

[9:03](https://www.youtube.com/watch?v=REascnFlq_8&t=543s) And all the agents have access to that

[9:04](https://www.youtube.com/watch?v=REascnFlq_8&t=544s) stuff. Because they have access to it,

[9:07](https://www.youtube.com/watch?v=REascnFlq_8&t=547s) you no longer have information that

[9:08](https://www.youtube.com/watch?v=REascnFlq_8&t=548s) would be okay to be shared that's stuck

[9:10](https://www.youtube.com/watch?v=REascnFlq_8&t=550s) in a silo. It now automatically filters

[9:12](https://www.youtube.com/watch?v=REascnFlq_8&t=552s) out into this public space. So,

[9:14](https://www.youtube.com/watch?v=REascnFlq_8&t=554s) examples, shared skills. If you code in

[9:17](https://www.youtube.com/watch?v=REascnFlq_8&t=557s) an organization, probably in your repo

[9:19](https://www.youtube.com/watch?v=REascnFlq_8&t=559s) you have shared skills. Anyone can make

[9:21](https://www.youtube.com/watch?v=REascnFlq_8&t=561s) them better. Someone has a better way

[9:22](https://www.youtube.com/watch?v=REascnFlq_8&t=562s) to, you know, profile your database or

[9:24](https://www.youtube.com/watch?v=REascnFlq_8&t=564s) whatever. They can write the skill. Next

[9:26](https://www.youtube.com/watch?v=REascnFlq_8&t=566s) time someone's sitting there is like,

[9:27](https://www.youtube.com/watch?v=REascnFlq_8&t=567s) "Oh my god, the database query is slow."

[9:29](https://www.youtube.com/watch?v=REascnFlq_8&t=569s) It uses the

[9:31](https://www.youtube.com/watch?v=REascnFlq_8&t=571s) profiling skill and everyone's a better

[9:32](https://www.youtube.com/watch?v=REascnFlq_8&t=572s) engineer. So, that's one version. The

[9:34](https://www.youtube.com/watch?v=REascnFlq_8&t=574s) other one that's pretty popular is

[9:35](https://www.youtube.com/watch?v=REascnFlq_8&t=575s) people decide they have some shared

[9:37](https://www.youtube.com/watch?v=REascnFlq_8&t=577s) mediums, like a wiki, airtable, etc. And

[9:40](https://www.youtube.com/watch?v=REascnFlq_8&t=580s) they uh they have a scale that says,

[9:42](https://www.youtube.com/watch?v=REascnFlq_8&t=582s) "Hey, put more data in there over time."

[9:44](https://www.youtube.com/watch?v=REascnFlq_8&t=584s) So, these are cool.

[9:45](https://www.youtube.com/watch?v=REascnFlq_8&t=585s) Um and they work as long as your your

[9:48](https://www.youtube.com/watch?v=REascnFlq_8&t=588s) agents have those tools and also some

[9:51](https://www.youtube.com/watch?v=REascnFlq_8&t=591s) trajectory incentives to really like get

[9:53](https://www.youtube.com/watch?v=REascnFlq_8&t=593s) data out in and out of of these shared

[9:55](https://www.youtube.com/watch?v=REascnFlq_8&t=595s) silos. Um

[9:57](https://www.youtube.com/watch?v=REascnFlq_8&t=597s) I think the next version of this that a

[9:59](https://www.youtube.com/watch?v=REascnFlq_8&t=599s) few people are working on is like you

[10:01](https://www.youtube.com/watch?v=REascnFlq_8&t=601s) have a sweeper AI. So, this actually if

[10:03](https://www.youtube.com/watch?v=REascnFlq_8&t=603s) there's one good idea in this talk

[10:06](https://www.youtube.com/watch?v=REascnFlq_8&t=606s) that I think works really well is this.

[10:08](https://www.youtube.com/watch?v=REascnFlq_8&t=608s) It's a sweeper AI. So, you have an AI

[10:10](https://www.youtube.com/watch?v=REascnFlq_8&t=610s) inside each private silo.

[10:12](https://www.youtube.com/watch?v=REascnFlq_8&t=612s) An AI has a policy about what has to

[10:15](https://www.youtube.com/watch?v=REascnFlq_8&t=615s) stay in the silo.

[10:17](https://www.youtube.com/watch?v=REascnFlq_8&t=617s) And then it also has a description of

[10:20](https://www.youtube.com/watch?v=REascnFlq_8&t=620s) all the shared spaces that you have. And

[10:22](https://www.youtube.com/watch?v=REascnFlq_8&t=622s) at the end of the day, it looks at new

[10:23](https://www.youtube.com/watch?v=REascnFlq_8&t=623s) information in the silo

[10:25](https://www.youtube.com/watch?v=REascnFlq_8&t=625s) and it puts it in the public spaces.

[10:26](https://www.youtube.com/watch?v=REascnFlq_8&t=626s) Well, public, public to your company.

[10:28](https://www.youtube.com/watch?v=REascnFlq_8&t=628s) So, this is the same as the personal

[10:30](https://www.youtube.com/watch?v=REascnFlq_8&t=630s) wiki that you all have AI building for

[10:32](https://www.youtube.com/watch?v=REascnFlq_8&t=632s) you at the end of the day so that I can

[10:33](https://www.youtube.com/watch?v=REascnFlq_8&t=633s) know your goals and your friends and all

[10:34](https://www.youtube.com/watch?v=REascnFlq_8&t=634s) that stuff, but it's at the company

[10:36](https://www.youtube.com/watch?v=REascnFlq_8&t=636s) level.

[10:37](https://www.youtube.com/watch?v=REascnFlq_8&t=637s) Um

[10:38](https://www.youtube.com/watch?v=REascnFlq_8&t=638s) the hard part is how do we pick what

[10:39](https://www.youtube.com/watch?v=REascnFlq_8&t=639s) private information is okay to to share

[10:42](https://www.youtube.com/watch?v=REascnFlq_8&t=642s) and to put it in shared silos. And I

[10:44](https://www.youtube.com/watch?v=REascnFlq_8&t=644s) think there's two approaches. There's

[10:46](https://www.youtube.com/watch?v=REascnFlq_8&t=646s) the ask a human approach. So, this is

[10:48](https://www.youtube.com/watch?v=REascnFlq_8&t=648s) like

[10:50](https://www.youtube.com/watch?v=REascnFlq_8&t=650s) the LLM comes up with a list of things

[10:51](https://www.youtube.com/watch?v=REascnFlq_8&t=651s) to contribute and then it asks the user,

[10:54](https://www.youtube.com/watch?v=REascnFlq_8&t=654s) "Hey, are you okay with me putting this

[10:55](https://www.youtube.com/watch?v=REascnFlq_8&t=655s) in the shared space?"

[10:56](https://www.youtube.com/watch?v=REascnFlq_8&t=656s) And you read it. You're like, "Yep."

[10:58](https://www.youtube.com/watch?v=REascnFlq_8&t=658s) Saved you a bunch of time.

[10:59](https://www.youtube.com/watch?v=REascnFlq_8&t=659s) Right? I mean, you weren't going to do

[11:00](https://www.youtube.com/watch?v=REascnFlq_8&t=660s) it otherwise.

[11:02](https://www.youtube.com/watch?v=REascnFlq_8&t=662s) Uh

[11:02](https://www.youtube.com/watch?v=REascnFlq_8&t=662s) I think the other version is you

[11:03](https://www.youtube.com/watch?v=REascnFlq_8&t=663s) actually ask the LLM to enforce a

[11:06](https://www.youtube.com/watch?v=REascnFlq_8&t=666s) policy.

[11:07](https://www.youtube.com/watch?v=REascnFlq_8&t=667s) And I think that actually is where

[11:09](https://www.youtube.com/watch?v=REascnFlq_8&t=669s) things are going to go very, very

[11:10](https://www.youtube.com/watch?v=REascnFlq_8&t=670s) quickly. Um and I think in the next 6

[11:13](https://www.youtube.com/watch?v=REascnFlq_8&t=673s) months we'll have a bunch of systems

[11:14](https://www.youtube.com/watch?v=REascnFlq_8&t=674s) where companies have trusted an LLM with

[11:16](https://www.youtube.com/watch?v=REascnFlq_8&t=676s) a policy to automatically surface more

[11:19](https://www.youtube.com/watch?v=REascnFlq_8&t=679s) and more information that otherwise

[11:20](https://www.youtube.com/watch?v=REascnFlq_8&t=680s) would have been private into a public

[11:22](https://www.youtube.com/watch?v=REascnFlq_8&t=682s) space. If you're like at a Fortune 500

[11:24](https://www.youtube.com/watch?v=REascnFlq_8&t=684s) enterprise company, unfortunately,

[11:26](https://www.youtube.com/watch?v=REascnFlq_8&t=686s) I don't think that's going to happen for

[11:27](https://www.youtube.com/watch?v=REascnFlq_8&t=687s) a while, but I think if you look at

[11:28](https://www.youtube.com/watch?v=REascnFlq_8&t=688s) smaller companies, like 10, 50 person

[11:30](https://www.youtube.com/watch?v=REascnFlq_8&t=690s) employees, high trust, like low

[11:32](https://www.youtube.com/watch?v=REascnFlq_8&t=692s) likelihood of something one one doing

[11:34](https://www.youtube.com/watch?v=REascnFlq_8&t=694s) bad with the data, which really clear to

[11:36](https://www.youtube.com/watch?v=REascnFlq_8&t=696s) know what data couldn't be shared,

[11:37](https://www.youtube.com/watch?v=REascnFlq_8&t=697s) basically finance and HR data, you're

[11:40](https://www.youtube.com/watch?v=REascnFlq_8&t=700s) going to see a ton of this. And the cool

[11:42](https://www.youtube.com/watch?v=REascnFlq_8&t=702s) thing here is this really improves

[11:43](https://www.youtube.com/watch?v=REascnFlq_8&t=703s) trajectories of systems on on common

[11:46](https://www.youtube.com/watch?v=REascnFlq_8&t=706s) work.

[11:47](https://www.youtube.com/watch?v=REascnFlq_8&t=707s) All right, that was third approach.

[11:49](https://www.youtube.com/watch?v=REascnFlq_8&t=709s) Fourth approach pretty obvious, use

[11:50](https://www.youtube.com/watch?v=REascnFlq_8&t=710s) humans as the conduit for information.

[11:52](https://www.youtube.com/watch?v=REascnFlq_8&t=712s) So this is like traditional agent to

[11:54](https://www.youtube.com/watch?v=REascnFlq_8&t=714s) agent.

[11:55](https://www.youtube.com/watch?v=REascnFlq_8&t=715s) My agent ask your agent, "Hey,

[11:58](https://www.youtube.com/watch?v=REascnFlq_8&t=718s) who is connected to someone on the

[12:01](https://www.youtube.com/watch?v=REascnFlq_8&t=721s) finance team at Acme Corp?"

[12:03](https://www.youtube.com/watch?v=REascnFlq_8&t=723s) You as a human see the request and

[12:05](https://www.youtube.com/watch?v=REascnFlq_8&t=725s) you're like, "Yeah, I'm okay with that.

[12:07](https://www.youtube.com/watch?v=REascnFlq_8&t=727s) Go and find the information inside of my

[12:09](https://www.youtube.com/watch?v=REascnFlq_8&t=729s) email."

[12:10](https://www.youtube.com/watch?v=REascnFlq_8&t=730s) And then it shows you the result. And

[12:12](https://www.youtube.com/watch?v=REascnFlq_8&t=732s) then you're like, "Yes, I'm okay with

[12:13](https://www.youtube.com/watch?v=REascnFlq_8&t=733s) that result

[12:14](https://www.youtube.com/watch?v=REascnFlq_8&t=734s) going to the person who asked." The big

[12:17](https://www.youtube.com/watch?v=REascnFlq_8&t=737s) problem with it is for for any request

[12:19](https://www.youtube.com/watch?v=REascnFlq_8&t=739s) that has low where it's like only a few

[12:22](https://www.youtube.com/watch?v=REascnFlq_8&t=742s) people will have the information, you're

[12:23](https://www.youtube.com/watch?v=REascnFlq_8&t=743s) kind of spamming everyone the request.

[12:25](https://www.youtube.com/watch?v=REascnFlq_8&t=745s) So if I ask this question to 100-person

[12:27](https://www.youtube.com/watch?v=REascnFlq_8&t=747s) company, 100 people are being pinged on

[12:29](https://www.youtube.com/watch?v=REascnFlq_8&t=749s) Slack, being like approves on these

[12:30](https://www.youtube.com/watch?v=REascnFlq_8&t=750s) requests to like farm your personal

[12:33](https://www.youtube.com/watch?v=REascnFlq_8&t=753s) network for this for this, you know, for

[12:35](https://www.youtube.com/watch?v=REascnFlq_8&t=755s) this like the answer to this question.

[12:38](https://www.youtube.com/watch?v=REascnFlq_8&t=758s) That's not very efficient. Um

[12:40](https://www.youtube.com/watch?v=REascnFlq_8&t=760s) And so that's why there's a better

[12:41](https://www.youtube.com/watch?v=REascnFlq_8&t=761s) version of it.

[12:43](https://www.youtube.com/watch?v=REascnFlq_8&t=763s) Um which I this is this is very

[12:45](https://www.youtube.com/watch?v=REascnFlq_8&t=765s) powerful, but uh I haven't seen it in

[12:47](https://www.youtube.com/watch?v=REascnFlq_8&t=767s) practice much. It's it's a black box

[12:49](https://www.youtube.com/watch?v=REascnFlq_8&t=769s) approach.

[12:51](https://www.youtube.com/watch?v=REascnFlq_8&t=771s) I wish I had a diagram for this.

[12:52](https://www.youtube.com/watch?v=REascnFlq_8&t=772s) Unfortunately for you all, I do not. So

[12:54](https://www.youtube.com/watch?v=REascnFlq_8&t=774s) here's what this means. The black box

[12:56](https://www.youtube.com/watch?v=REascnFlq_8&t=776s) approach is where when you ask a

[12:58](https://www.youtube.com/watch?v=REascnFlq_8&t=778s) question

[13:00](https://www.youtube.com/watch?v=REascnFlq_8&t=780s) that can only be answered by looking at

[13:02](https://www.youtube.com/watch?v=REascnFlq_8&t=782s) information in other people's silos. You

[13:04](https://www.youtube.com/watch?v=REascnFlq_8&t=784s) have an LLM,

[13:06](https://www.youtube.com/watch?v=REascnFlq_8&t=786s) the trace of which no one has access to,

[13:08](https://www.youtube.com/watch?v=REascnFlq_8&t=788s) that gets access to all of the data,

[13:11](https://www.youtube.com/watch?v=REascnFlq_8&t=791s) and it gets to the answer.

[13:13](https://www.youtube.com/watch?v=REascnFlq_8&t=793s) Right? Like by say get to the answer,

[13:15](https://www.youtube.com/watch?v=REascnFlq_8&t=795s) either gets the answer or it's about to

[13:17](https://www.youtube.com/watch?v=REascnFlq_8&t=797s) do the the the any tool call that's all

[13:19](https://www.youtube.com/watch?v=REascnFlq_8&t=799s) right.

[13:20](https://www.youtube.com/watch?v=REascnFlq_8&t=800s) And then it looks at what information

[13:22](https://www.youtube.com/watch?v=REascnFlq_8&t=802s) did it need to make that tool call,

[13:24](https://www.youtube.com/watch?v=REascnFlq_8&t=804s) and it only ask the people

[13:26](https://www.youtube.com/watch?v=REascnFlq_8&t=806s) who own that information for their

[13:28](https://www.youtube.com/watch?v=REascnFlq_8&t=808s) approval to do the tool call.

[13:30](https://www.youtube.com/watch?v=REascnFlq_8&t=810s) So, in the example before that I gave,

[13:31](https://www.youtube.com/watch?v=REascnFlq_8&t=811s) when I ask 100 people at my company,

[13:33](https://www.youtube.com/watch?v=REascnFlq_8&t=813s) "Hey, do you know the CFO at Acme Corp?"

[13:36](https://www.youtube.com/watch?v=REascnFlq_8&t=816s) The request goes to everyone's agents in

[13:39](https://www.youtube.com/watch?v=REascnFlq_8&t=819s) my company.

[13:41](https://www.youtube.com/watch?v=REascnFlq_8&t=821s) All of their agents look in their Gmail

[13:43](https://www.youtube.com/watch?v=REascnFlq_8&t=823s) and their private silos to see if

[13:45](https://www.youtube.com/watch?v=REascnFlq_8&t=825s) they're connected to the CFO.

[13:47](https://www.youtube.com/watch?v=REascnFlq_8&t=827s) That happens automatically. No No human

[13:49](https://www.youtube.com/watch?v=REascnFlq_8&t=829s) is being asked for approval for that to

[13:50](https://www.youtube.com/watch?v=REascnFlq_8&t=830s) happen.

[13:51](https://www.youtube.com/watch?v=REascnFlq_8&t=831s) Then it

[13:53](https://www.youtube.com/watch?v=REascnFlq_8&t=833s) the agent in the black box ha gets the

[13:55](https://www.youtube.com/watch?v=REascnFlq_8&t=835s) list of the 20 people who are connected.

[13:59](https://www.youtube.com/watch?v=REascnFlq_8&t=839s) It looks at contacts from the emails to

[14:01](https://www.youtube.com/watch?v=REascnFlq_8&t=841s) determine who has the strongest

[14:02](https://www.youtube.com/watch?v=REascnFlq_8&t=842s) connection. It determines that it's Bob.

[14:06](https://www.youtube.com/watch?v=REascnFlq_8&t=846s) And then, it just asks Bob, "Hey,

[14:09](https://www.youtube.com/watch?v=REascnFlq_8&t=849s) Jean-Denis

[14:10](https://www.youtube.com/watch?v=REascnFlq_8&t=850s) wants you to introduce him

[14:12](https://www.youtube.com/watch?v=REascnFlq_8&t=852s) to Jane, the CFO at Acme Corp.

[14:14](https://www.youtube.com/watch?v=REascnFlq_8&t=854s) I know you're well connected to her. Am

[14:16](https://www.youtube.com/watch?v=REascnFlq_8&t=856s) I okay sharing that bit of information

[14:17](https://www.youtube.com/watch?v=REascnFlq_8&t=857s) with Jean-Denis?"

[14:19](https://www.youtube.com/watch?v=REascnFlq_8&t=859s) And you're like, "Yeah, sure." You click

[14:20](https://www.youtube.com/watch?v=REascnFlq_8&t=860s) yes. No big deal.

[14:21](https://www.youtube.com/watch?v=REascnFlq_8&t=861s) The important thing is you have to trust

[14:23](https://www.youtube.com/watch?v=REascnFlq_8&t=863s) the black box. So, you have to trust

[14:24](https://www.youtube.com/watch?v=REascnFlq_8&t=864s) that you can break down all the silos

[14:28](https://www.youtube.com/watch?v=REascnFlq_8&t=868s) for an LLM that has full access

[14:31](https://www.youtube.com/watch?v=REascnFlq_8&t=871s) and that doesn't ask for permission

[14:33](https://www.youtube.com/watch?v=REascnFlq_8&t=873s) until there's this sharing moment or

[14:35](https://www.youtube.com/watch?v=REascnFlq_8&t=875s) this right step.

[14:37](https://www.youtube.com/watch?v=REascnFlq_8&t=877s) So, actually, within a company, this is

[14:38](https://www.youtube.com/watch?v=REascnFlq_8&t=878s) not impossible to do.

[14:40](https://www.youtube.com/watch?v=REascnFlq_8&t=880s) Uh

[14:41](https://www.youtube.com/watch?v=REascnFlq_8&t=881s) and actually, your security and

[14:43](https://www.youtube.com/watch?v=REascnFlq_8&t=883s) compliance team can get okay with it. Um

[14:46](https://www.youtube.com/watch?v=REascnFlq_8&t=886s) you just have to have You have to be

[14:47](https://www.youtube.com/watch?v=REascnFlq_8&t=887s) sure that the human in the loop step is

[14:50](https://www.youtube.com/watch?v=REascnFlq_8&t=890s) correct, and you have to be sure that

[14:52](https://www.youtube.com/watch?v=REascnFlq_8&t=892s) you're not letting other information go

[14:54](https://www.youtube.com/watch?v=REascnFlq_8&t=894s) through without last answer. So, you

[14:56](https://www.youtube.com/watch?v=REascnFlq_8&t=896s) know, like the the nightmare scenarios

[14:58](https://www.youtube.com/watch?v=REascnFlq_8&t=898s) and things like this are things like uh

[15:01](https://www.youtube.com/watch?v=REascnFlq_8&t=901s) um

[15:02](https://www.youtube.com/watch?v=REascnFlq_8&t=902s) Sorry. I'm like

[15:05](https://www.youtube.com/watch?v=REascnFlq_8&t=905s) We have plenty of time. I'm almost done.

[15:06](https://www.youtube.com/watch?v=REascnFlq_8&t=906s) So, it's great. Um the nightmare

[15:08](https://www.youtube.com/watch?v=REascnFlq_8&t=908s) scenarios with things like this is

[15:09](https://www.youtube.com/watch?v=REascnFlq_8&t=909s) someone asks a question like,

[15:11](https://www.youtube.com/watch?v=REascnFlq_8&t=911s) "Are you connected to a recruiter at the

[15:13](https://www.youtube.com/watch?v=REascnFlq_8&t=913s) other company that you have no business

[15:15](https://www.youtube.com/watch?v=REascnFlq_8&t=915s) being recruited to as a way for them to

[15:17](https://www.youtube.com/watch?v=REascnFlq_8&t=917s) find out that you're interviewing

[15:18](https://www.youtube.com/watch?v=REascnFlq_8&t=918s) somewhere else, right?" So, you know,

[15:20](https://www.youtube.com/watch?v=REascnFlq_8&t=920s) there are you you still sometimes with a

[15:22](https://www.youtube.com/watch?v=REascnFlq_8&t=922s) black box inadvertently get an

[15:24](https://www.youtube.com/watch?v=REascnFlq_8&t=924s) information out that you shouldn't be

[15:26](https://www.youtube.com/watch?v=REascnFlq_8&t=926s) able to. You have to really think about

[15:28](https://www.youtube.com/watch?v=REascnFlq_8&t=928s) how you would build a great system.

[15:30](https://www.youtube.com/watch?v=REascnFlq_8&t=930s) So, those are the those are the

[15:31](https://www.youtube.com/watch?v=REascnFlq_8&t=931s) approaches. I think if I were to bet on

[15:34](https://www.youtube.com/watch?v=REascnFlq_8&t=934s) one that has immediate ROI that we're

[15:36](https://www.youtube.com/watch?v=REascnFlq_8&t=936s) going to all see in both like open

[15:38](https://www.youtube.com/watch?v=REascnFlq_8&t=938s) source claw-ish worlds and then like

[15:41](https://www.youtube.com/watch?v=REascnFlq_8&t=941s) small companies, it's going to be the

[15:42](https://www.youtube.com/watch?v=REascnFlq_8&t=942s) wiki that's automatically created by AI

[15:44](https://www.youtube.com/watch?v=REascnFlq_8&t=944s) like the information base that's kept up

[15:46](https://www.youtube.com/watch?v=REascnFlq_8&t=946s) to date. I think there will be database

[15:47](https://www.youtube.com/watch?v=REascnFlq_8&t=947s) versions of it, wiki versions of it, and

[15:50](https://www.youtube.com/watch?v=REascnFlq_8&t=950s) I think more and more we're going to

[15:51](https://www.youtube.com/watch?v=REascnFlq_8&t=951s) trust LLMs to make the decision about

[15:53](https://www.youtube.com/watch?v=REascnFlq_8&t=953s) what's okay to share and what's not.

[15:55](https://www.youtube.com/watch?v=REascnFlq_8&t=955s) There are problems. So, prompt injection

[15:58](https://www.youtube.com/watch?v=REascnFlq_8&t=958s) in in the silos can be a real problem.

[16:01](https://www.youtube.com/watch?v=REascnFlq_8&t=961s) Obviously. So, if you have a silo that

[16:03](https://www.youtube.com/watch?v=REascnFlq_8&t=963s) has like that's more open and someone

[16:05](https://www.youtube.com/watch?v=REascnFlq_8&t=965s) can put something bad in there and then

[16:07](https://www.youtube.com/watch?v=REascnFlq_8&t=967s) that as part of the energetic search you

[16:09](https://www.youtube.com/watch?v=REascnFlq_8&t=969s) pull it out, you know, bad things can

[16:11](https://www.youtube.com/watch?v=REascnFlq_8&t=971s) can happen.

[16:12](https://www.youtube.com/watch?v=REascnFlq_8&t=972s) Uh you can have it's very easy to have a

[16:14](https://www.youtube.com/watch?v=REascnFlq_8&t=974s) shared wiki that just goes like totally

[16:16](https://www.youtube.com/watch?v=REascnFlq_8&t=976s) off the rails,

[16:17](https://www.youtube.com/watch?v=REascnFlq_8&t=977s) you know, like the information there one

[16:19](https://www.youtube.com/watch?v=REascnFlq_8&t=979s) piece of information there is incorrect

[16:21](https://www.youtube.com/watch?v=REascnFlq_8&t=981s) cuz LLM made a mistake and then it like

[16:23](https://www.youtube.com/watch?v=REascnFlq_8&t=983s) poisons it forever. I have a personal

[16:25](https://www.youtube.com/watch?v=REascnFlq_8&t=985s) wiki that thinks my agent's name is Apex

[16:27](https://www.youtube.com/watch?v=REascnFlq_8&t=987s) right now, but I renamed my agent a

[16:29](https://www.youtube.com/watch?v=REascnFlq_8&t=989s) month ago to Ivy. And like somewhere in

[16:31](https://www.youtube.com/watch?v=REascnFlq_8&t=991s) memory bank uh uh

[16:33](https://www.youtube.com/watch?v=REascnFlq_8&t=993s) of my like setup uh Apex lives and so I

[16:37](https://www.youtube.com/watch?v=REascnFlq_8&t=997s) can't get rid of it.

[16:38](https://www.youtube.com/watch?v=REascnFlq_8&t=998s) Uh that's fine for Apex. That's a funny

[16:40](https://www.youtube.com/watch?v=REascnFlq_8&t=1000s) one, but it's like much more difficult

[16:41](https://www.youtube.com/watch?v=REascnFlq_8&t=1001s) if it's a really wrong piece of

[16:43](https://www.youtube.com/watch?v=REascnFlq_8&t=1003s) information about your business.

[16:45](https://www.youtube.com/watch?v=REascnFlq_8&t=1005s) Um

[16:46](https://www.youtube.com/watch?v=REascnFlq_8&t=1006s) if you don't have human in the loop for

[16:48](https://www.youtube.com/watch?v=REascnFlq_8&t=1008s) any of the steps, obviously there'll be

[16:50](https://www.youtube.com/watch?v=REascnFlq_8&t=1010s) false and wrong disclosures. You know,

[16:53](https://www.youtube.com/watch?v=REascnFlq_8&t=1013s) sometimes when there's a wrong

[16:54](https://www.youtube.com/watch?v=REascnFlq_8&t=1014s) disclosure of information, someone gets

[16:56](https://www.youtube.com/watch?v=REascnFlq_8&t=1016s) fired. Someone there's a wrong

[16:58](https://www.youtube.com/watch?v=REascnFlq_8&t=1018s) disclosure,

[16:59](https://www.youtube.com/watch?v=REascnFlq_8&t=1019s) it doesn't matter at all. Sometimes a

[17:01](https://www.youtube.com/watch?v=REascnFlq_8&t=1021s) customer sues you. So, you know, uh you

[17:03](https://www.youtube.com/watch?v=REascnFlq_8&t=1023s) got to be careful.

[17:05](https://www.youtube.com/watch?v=REascnFlq_8&t=1025s) Um and then I think

[17:08](https://www.youtube.com/watch?v=REascnFlq_8&t=1028s) this all sounds nice, but like who

[17:09](https://www.youtube.com/watch?v=REascnFlq_8&t=1029s) approves what, what's logged, what's

[17:11](https://www.youtube.com/watch?v=REascnFlq_8&t=1031s) reversible? The black box idea is really

[17:13](https://www.youtube.com/watch?v=REascnFlq_8&t=1033s) great, but it can't truly be a black

[17:15](https://www.youtube.com/watch?v=REascnFlq_8&t=1035s) box. Someone at your company will want

[17:17](https://www.youtube.com/watch?v=REascnFlq_8&t=1037s) to audit it at some point. They want to

[17:19](https://www.youtube.com/watch?v=REascnFlq_8&t=1039s) understand what's going in there, right?

[17:20](https://www.youtube.com/watch?v=REascnFlq_8&t=1040s) So, at some level there must be some

[17:22](https://www.youtube.com/watch?v=REascnFlq_8&t=1042s) person in the CISO suite or somewhere

[17:24](https://www.youtube.com/watch?v=REascnFlq_8&t=1044s) that has access to all the data.

[17:26](https://www.youtube.com/watch?v=REascnFlq_8&t=1046s) Um

[17:28](https://www.youtube.com/watch?v=REascnFlq_8&t=1048s) Yeah.

[17:29](https://www.youtube.com/watch?v=REascnFlq_8&t=1049s) So,

[17:30](https://www.youtube.com/watch?v=REascnFlq_8&t=1050s) um what do I think? Well, I do think the

[17:32](https://www.youtube.com/watch?v=REascnFlq_8&t=1052s) frontier is auto. So, I've said that. I

[17:34](https://www.youtube.com/watch?v=REascnFlq_8&t=1054s) think in coding we used to approve

[17:36](https://www.youtube.com/watch?v=REascnFlq_8&t=1056s) everything. Then we were like, "YOLO,

[17:37](https://www.youtube.com/watch?v=REascnFlq_8&t=1057s) live dangerously." And now the gods at

[17:39](https://www.youtube.com/watch?v=REascnFlq_8&t=1059s) Anthropic have granted us auto mode. And

[17:42](https://www.youtube.com/watch?v=REascnFlq_8&t=1062s) auto mode tries to figure out when we're

[17:43](https://www.youtube.com/watch?v=REascnFlq_8&t=1063s) maybe being a little silly and it tells

[17:45](https://www.youtube.com/watch?v=REascnFlq_8&t=1065s) us. Well, I think A to A across

[17:48](https://www.youtube.com/watch?v=REascnFlq_8&t=1068s) information silos will be the same way.

[17:50](https://www.youtube.com/watch?v=REascnFlq_8&t=1070s) I think what's going to happen is we're

[17:51](https://www.youtube.com/watch?v=REascnFlq_8&t=1071s) going to get comfortable with low

[17:52](https://www.youtube.com/watch?v=REascnFlq_8&t=1072s) sensitivity information being pulled out

[17:54](https://www.youtube.com/watch?v=REascnFlq_8&t=1074s) and put into common spaces. And then we

[17:56](https://www.youtube.com/watch?v=REascnFlq_8&t=1076s) will have a place that's like human

[17:58](https://www.youtube.com/watch?v=REascnFlq_8&t=1078s) review or always human approved.

[18:00](https://www.youtube.com/watch?v=REascnFlq_8&t=1080s) And then over time what's going to

[18:02](https://www.youtube.com/watch?v=REascnFlq_8&t=1082s) happen is the LLMs will get more

[18:03](https://www.youtube.com/watch?v=REascnFlq_8&t=1083s) powerful.

[18:04](https://www.youtube.com/watch?v=REascnFlq_8&t=1084s) We will be better at encoding safe

[18:06](https://www.youtube.com/watch?v=REascnFlq_8&t=1086s) policies within them. We'll be better at

[18:08](https://www.youtube.com/watch?v=REascnFlq_8&t=1088s) designing for the really hard areas

[18:10](https://www.youtube.com/watch?v=REascnFlq_8&t=1090s) tools that get the privacy trade-off

[18:12](https://www.youtube.com/watch?v=REascnFlq_8&t=1092s) correct.

[18:13](https://www.youtube.com/watch?v=REascnFlq_8&t=1093s) And it'll just be more and more auto for

[18:15](https://www.youtube.com/watch?v=REascnFlq_8&t=1095s) building shared silos and even sometimes

[18:17](https://www.youtube.com/watch?v=REascnFlq_8&t=1097s) for deciding whether involve a human.

[18:19](https://www.youtube.com/watch?v=REascnFlq_8&t=1099s) So, the example that I have is like if I

[18:22](https://www.youtube.com/watch?v=REascnFlq_8&t=1102s) ask for notes from a weekly recurring

[18:24](https://www.youtube.com/watch?v=REascnFlq_8&t=1104s) call with, you know, a supplier of ours

[18:27](https://www.youtube.com/watch?v=REascnFlq_8&t=1107s) and their finance team, maybe the the

[18:29](https://www.youtube.com/watch?v=REascnFlq_8&t=1109s) LLM's like, "Oh, well, in giving your

[18:31](https://www.youtube.com/watch?v=REascnFlq_8&t=1111s) role, I don't need to ask anyone on

[18:33](https://www.youtube.com/watch?v=REascnFlq_8&t=1113s) those teams for permission. I can just

[18:34](https://www.youtube.com/watch?v=REascnFlq_8&t=1114s) share the notes with you. It's like it's

[18:36](https://www.youtube.com/watch?v=REascnFlq_8&t=1116s) fine. I they can it can look at the

[18:38](https://www.youtube.com/watch?v=REascnFlq_8&t=1118s) content. It can see what my role is. It

[18:39](https://www.youtube.com/watch?v=REascnFlq_8&t=1119s) can decide from a risk perspective I'm

[18:42](https://www.youtube.com/watch?v=REascnFlq_8&t=1122s) okay with that disclosure."

[18:43](https://www.youtube.com/watch?v=REascnFlq_8&t=1123s) Um and the cool thing about auto, by the

[18:46](https://www.youtube.com/watch?v=REascnFlq_8&t=1126s) way, is if you design your systems that

[18:47](https://www.youtube.com/watch?v=REascnFlq_8&t=1127s) way, it'll scale with model capacity.

[18:49](https://www.youtube.com/watch?v=REascnFlq_8&t=1129s) So, my encouragement would be like, you

[18:51](https://www.youtube.com/watch?v=REascnFlq_8&t=1131s) need to start if you have agent to agent

[18:53](https://www.youtube.com/watch?v=REascnFlq_8&t=1133s) or or work across silos, which is I

[18:55](https://www.youtube.com/watch?v=REascnFlq_8&t=1135s) think the better way to think about it,

[18:56](https://www.youtube.com/watch?v=REascnFlq_8&t=1136s) definitely define a low sensitivity zone

[18:59](https://www.youtube.com/watch?v=REascnFlq_8&t=1139s) where you're okay with the LLM making a

[19:00](https://www.youtube.com/watch?v=REascnFlq_8&t=1140s) call.

[19:02](https://www.youtube.com/watch?v=REascnFlq_8&t=1142s) And get okay with that.

[19:04](https://www.youtube.com/watch?v=REascnFlq_8&t=1144s) And then magically, as time goes on,

[19:07](https://www.youtube.com/watch?v=REascnFlq_8&t=1147s) it'll get bigger and your system will

[19:08](https://www.youtube.com/watch?v=REascnFlq_8&t=1148s) naturally get more powerful, which is

[19:09](https://www.youtube.com/watch?v=REascnFlq_8&t=1149s) what you you So, you want to be on a

[19:11](https://www.youtube.com/watch?v=REascnFlq_8&t=1151s) beach.

[19:13](https://www.youtube.com/watch?v=REascnFlq_8&t=1153s) That's what you want to do. That's where

[19:14](https://www.youtube.com/watch?v=REascnFlq_8&t=1154s) I want to be. My kids

[19:16](https://www.youtube.com/watch?v=REascnFlq_8&t=1156s) in Hawaii. Okay, network effects. I've 1

[19:18](https://www.youtube.com/watch?v=REascnFlq_8&t=1158s) minute.

[19:19](https://www.youtube.com/watch?v=REascnFlq_8&t=1159s) Uh this is the conclusion. So, we talked

[19:22](https://www.youtube.com/watch?v=REascnFlq_8&t=1162s) about five approaches, blah blah blah,

[19:24](https://www.youtube.com/watch?v=REascnFlq_8&t=1164s) trust boundaries, custom tools, shared

[19:25](https://www.youtube.com/watch?v=REascnFlq_8&t=1165s) silos, humans in the loop, and this

[19:27](https://www.youtube.com/watch?v=REascnFlq_8&t=1167s) human in the black box version. Uh I

[19:30](https://www.youtube.com/watch?v=REascnFlq_8&t=1170s) think this stuff is very powerful.

[19:32](https://www.youtube.com/watch?v=REascnFlq_8&t=1172s) I think the interesting questions a

[19:33](https://www.youtube.com/watch?v=REascnFlq_8&t=1173s) little bit are

[19:35](https://www.youtube.com/watch?v=REascnFlq_8&t=1175s) in within companies, I think this will

[19:36](https://www.youtube.com/watch?v=REascnFlq_8&t=1176s) all work very soon. The big question is

[19:38](https://www.youtube.com/watch?v=REascnFlq_8&t=1178s) where how how how comfortable are we are

[19:40](https://www.youtube.com/watch?v=REascnFlq_8&t=1180s) we with something like an auto mode

[19:42](https://www.youtube.com/watch?v=REascnFlq_8&t=1182s) around privacy?

[19:43](https://www.youtube.com/watch?v=REascnFlq_8&t=1183s) And I think the really interesting

[19:44](https://www.youtube.com/watch?v=REascnFlq_8&t=1184s) question that I don't have an answer

[19:46](https://www.youtube.com/watch?v=REascnFlq_8&t=1186s) for, but I think whoever does this will

[19:47](https://www.youtube.com/watch?v=REascnFlq_8&t=1187s) be wealthier than I am, is if you can

[19:50](https://www.youtube.com/watch?v=REascnFlq_8&t=1190s) think of

[19:52](https://www.youtube.com/watch?v=REascnFlq_8&t=1192s) scenarios where you can get multiple

[19:54](https://www.youtube.com/watch?v=REascnFlq_8&t=1194s) companies to agree to their information

[19:57](https://www.youtube.com/watch?v=REascnFlq_8&t=1197s) silos

[19:59](https://www.youtube.com/watch?v=REascnFlq_8&t=1199s) having a common agent working across

[20:00](https://www.youtube.com/watch?v=REascnFlq_8&t=1200s) them. So, there's like a company that I

[20:02](https://www.youtube.com/watch?v=REascnFlq_8&t=1202s) won't name in in in in somewhere in the

[20:05](https://www.youtube.com/watch?v=REascnFlq_8&t=1205s) world, uh working on like finance stuff

[20:07](https://www.youtube.com/watch?v=REascnFlq_8&t=1207s) where they have a lot of investment

[20:08](https://www.youtube.com/watch?v=REascnFlq_8&t=1208s) banks. And actually the investment banks

[20:10](https://www.youtube.com/watch?v=REascnFlq_8&t=1210s) there's a benefit to them sharing

[20:11](https://www.youtube.com/watch?v=REascnFlq_8&t=1211s) private data about private companies for

[20:13](https://www.youtube.com/watch?v=REascnFlq_8&t=1213s) purposes of things like lending.

[20:15](https://www.youtube.com/watch?v=REascnFlq_8&t=1215s) And they're starting to look in this

[20:16](https://www.youtube.com/watch?v=REascnFlq_8&t=1216s) direction. Where they're trusting each

[20:18](https://www.youtube.com/watch?v=REascnFlq_8&t=1218s) other's agents to be able to work across

[20:20](https://www.youtube.com/watch?v=REascnFlq_8&t=1220s) what before would have been private

[20:21](https://www.youtube.com/watch?v=REascnFlq_8&t=1221s) information, with agents deciding what

[20:23](https://www.youtube.com/watch?v=REascnFlq_8&t=1223s) can be accessed or not. And it's it's

[20:25](https://www.youtube.com/watch?v=REascnFlq_8&t=1225s) cool. It's very cool. And I think once

[20:27](https://www.youtube.com/watch?v=REascnFlq_8&t=1227s) you find some use cases across across

[20:30](https://www.youtube.com/watch?v=REascnFlq_8&t=1230s) companies, uh

[20:32](https://www.youtube.com/watch?v=REascnFlq_8&t=1232s) I think that'll be a really good

[20:33](https://www.youtube.com/watch?v=REascnFlq_8&t=1233s) beachhead to to move more in this

[20:35](https://www.youtube.com/watch?v=REascnFlq_8&t=1235s) direction.

[20:36](https://www.youtube.com/watch?v=REascnFlq_8&t=1236s) So,

[20:37](https://www.youtube.com/watch?v=REascnFlq_8&t=1237s) yeah.

[20:38](https://www.youtube.com/watch?v=REascnFlq_8&t=1238s) You know, as a human though, I ask

[20:39](https://www.youtube.com/watch?v=REascnFlq_8&t=1239s) myself,

[20:41](https://www.youtube.com/watch?v=REascnFlq_8&t=1241s) do I trust the future where agents make

[20:43](https://www.youtube.com/watch?v=REascnFlq_8&t=1243s) all the decisions about privacy? I don't

[20:45](https://www.youtube.com/watch?v=REascnFlq_8&t=1245s) know about that. I just think it's a it

[20:46](https://www.youtube.com/watch?v=REascnFlq_8&t=1246s) is for better or worse the direction

[20:48](https://www.youtube.com/watch?v=REascnFlq_8&t=1248s) things are going.

[20:50](https://www.youtube.com/watch?v=REascnFlq_8&t=1250s) And I just I'm like so good on time. So,

[20:52](https://www.youtube.com/watch?v=REascnFlq_8&t=1252s) I'm on time. Thank you for coming. I

[20:54](https://www.youtube.com/watch?v=REascnFlq_8&t=1254s) again, uh

[20:56](https://www.youtube.com/watch?v=REascnFlq_8&t=1256s) yeah. [applause] Thanks for being here,

[20:57](https://www.youtube.com/watch?v=REascnFlq_8&t=1257s) and

[21:13](https://www.youtube.com/watch?v=REascnFlq_8&t=1273s) >> [music]

[21:16](https://www.youtube.com/watch?v=REascnFlq_8&t=1276s) >> Mhm.
