# Your agents lack context: Here's how to fix "You're absolutely right!" — Brandon Waselnuk, Unblocked

Brandon Waselnuk, Unblocked

[Talk page](https://ai.engineer/talks/KcVkq5L-0f0-your-agents-lack-context-heres-fix-youre) · [Video](https://www.youtube.com/watch?v=KcVkq5L-0f0)

## Your Agents Lack Context: How to Fix “You’re Absolutely Right!”

Brandon Waselnuk of Unblocked explains why capable coding agents still produce locally plausible, unmergeable work—and how identity-aware retrieval, conflict resolution, permissions, and structured queries can give them the organizational context they lack.

### The agent starts as a brilliant new hire

[0:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=12s)

AI-generated code should feel like it came from someone who has worked on your team for years. That standard is demanding because experienced engineers have absorbed far more than source code. Rejected pull requests taught them local conventions. Meetings supplied product history. An overnight production incident connected a rollout procedure to its consequences. In effect, the engineer’s brain became a context engine through repeated exposure to how this company actually works. [1:13](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=73s)

![Recording frame at 96 seconds](https://ai.engineer/api/talk-essay-media/aecda71341250ea2f1e76228ac876775e6d982ab360daa297943d83f7f8f766f)

Recording frame at 96 seconds

A new terminal session has none of that history. The model may be intelligent, but it does not know how this organization ships, which architectural decisions still apply, or what happened the last time a similar change reached production. If that missing context sends the agent in the wrong direction, every later action builds on a bad starting point. [1:51](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=111s)

### Bad context gets more expensive as autonomy increases

[2:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=137s)

Autocomplete kept the cost of a context failure small. A suggestion appeared, an engineer judged it using local knowledge, and either pressed Tab or rejected it. Moving toward agentic IDEs, parallel workers, and background agents removes that immediate filter. The agent must recognize when it has reached a wall, ask the right questions, and produce code that fits a revenue-bearing brownfield system—not merely a greenfield demo. [2:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=137s)

![Recording frame at 142 seconds](https://ai.engineer/api/talk-essay-media/019ccc29bf3cb0f38840d8c71c5a15558d969d062675e687b58c98c6e330c7f4)

Recording frame at 142 seconds

What changes as the human moves farther out of the loop? The diagram follows the accumulating cost. A wrong autocomplete suggestion is discarded immediately; a longer agent run spends search tokens and rework on repeated corrections; parallel agents create more output to review; and a background agent needs machine-accessible context if it is expected to continue without asking a person. The relationship to notice is compounding: greater autonomy magnifies an early context error rather than merely repeating it. [3:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=197s)

The familiar failure is the correction loop: the agent announces that it finished, the engineer rejects the result, and several rounds of repair follow. Those rounds consume search tokens and human time. Parallel agents add a review tax because reviewers also need business logic and operating context to distinguish a plausible patch from a mergeable one. Fully unattended work raises the requirement again: the context source must remain queryable while the agent operates. [3:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=197s)

The compounding cost of missing context
As human supervision decreases, an early misunderstanding survives longer and creates more search, rework, and review.
- Autocomplete — A human immediately accepts or rejects one suggestion.
- Agent correction loop — Wrong assumptions generate repeated search and rework.
- Parallel agents — More generated work creates a review tax.
- Background agents — Unattended workers need context they can query themselves.
- Autocomplete → Agent correction loop (Less immediate filtering)
- Agent correction loop → Parallel agents (More work reaches review)
- Parallel agents → Background agents (Human leaves the loop)

### Two useful approaches that plateau

[4:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=251s)

Static project instructions can improve an agent, but they create what Waselnuk calls the **curated context trap**. A team writes Markdown files, gives agents a filesystem to search, and sees better results. Then the operational problems begin: the files must be distributed, they age like other documentation, and someone must decide which version of the organization’s practices deserves to be authoritative for everyone. [4:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=251s)

![Recording frame at 360 seconds](https://ai.engineer/api/talk-essay-media/a4517d21a25e2b3a8ac0bc4a6621eebd1df064f4c664df0a4cf53deafac2003b)

Recording frame at 360 seconds

**MCP access** solves a different problem: it gives an agent a path to another system. It does not ensure that the agent calls the tool, because tool and server descriptions influence selection. Even when the call happens, the agent may stop at the first plausible result. Waselnuk describes this as **satisfaction of search**: an architecture record says to use approach B, so the agent proceeds without finding last night’s Slack conversation saying the team has switched to A. [5:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=311s)

Access is therefore not understanding. A patch can compile while remaining operationally wrong because the agent cannot see what lies “below the waterline”: rollout procedures, feature-flag sequencing, incident history, and current decisions that have not reached an architecture document. The concrete consequence is code that passes a local technical threshold and still causes a production incident. [5:43](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=343s)

### What a context engine must do before returning an answer

[6:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=372s)

Consider the request, “I want to get auth stood up.” The words alone do not identify the right repositories, conventions, or experts. A context engine can begin with the requester: where they commit, which code they touch, and who reviews their work. That identity narrows the initial search, and the resulting relationships provide pivots into the rest of the organization’s information. [6:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=372s)

![Recording frame at 462 seconds](https://ai.engineer/api/talk-essay-media/a78948a4d974de5c3f2b18922b301015da9aed69ad943edd2180c5f7213c775a)

Recording frame at 462 seconds

The engine then has to perform several distinct jobs:

- **Unified system context:** Search across the relevant systems instead of treating one repository or document store as the whole organization.
- **Targeted retrieval:** Return a linked document quickly when the task is narrow, while allowing deeper research when discovery is necessary.
- **Conflict resolution:** Detect that an old architecture diagram and a recent executive Slack thread disagree, then determine which evidence should guide the task rather than silently choosing the first result.
- **Personalized relevance:** Use who is asking, where they work, and what they are working on to focus retrieval.
- **Token optimization:** Compress the research into useful context instead of sending every discovered artifact to the model.
- **Permission enforcement:** Preserve OAuth, SSO, and source permissions so information from a restricted project does not leak into another user’s answer. [6:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=402s) [8:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=529s)

Permission enforcement must happen during retrieval, not after an unrestricted answer has already been assembled. If the requester cannot access “secret project A,” material from that project must not enter the response. Human-facing answers and machine-to-machine responses also need different packaging: a person in Slack may benefit from explanation, while an agent needs a compact packet that avoids spending tokens on presentation. [6:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=402s)

What data movement turns scattered workplace records into task-specific context? The engine ingests engineering sources, including real-time incident-management signals, reasons across them using the six capabilities, and emits an answer shaped for the destination workflow. The diagram makes the critical middle layer visible: connecting sources directly to an agent skips identity, conflict, permission, and compression decisions—the very work that turns reachable information into usable context. [7:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=439s)

From organizational data to workflow-specific context
The engine does more than retrieve: it interprets, filters, compresses, and shapes evidence before delivery.
- Organizational sources — Engineering data and real-time operational signals enter from workplace systems.
- Identity and relevance — The requester’s work, commits, and relationships focus the search.
- Retrieval and conflict resolution — The engine supports fast lookup or deeper research and reconciles contradictory evidence.
- Permission filtering — The requester’s access limits what may enter the response.
- Token-optimized context — Research is compressed and formatted for the destination.
- Human or agent workflow — The answer reaches Slack, a coding agent, or another task-specific interface.
- Organizational sources → Identity and relevance (Ingest and focus)
- Identity and relevance → Retrieval and conflict resolution (Narrow the territory)
- Retrieval and conflict resolution → Permission filtering (Filter allowable evidence)
- Permission filtering → Token-optimized context (Shape a safe response)
- Token-optimized context → Human or agent workflow (Deliver in the required form)

### The same prompt, with less repeated discovery

[8:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=529s)

Unblocked ran the same prompt against the same model twice, once with context and once without. The run without the engine used about **21 million tokens**; the context-assisted run used **10.8 million** and finished about **two hours sooner**. Waselnuk attributes much of the difference to removing repeated exploratory greps at the start of agent sessions: the assisted run began “hydrated” with relevant organizational context instead of rediscovering it. [9:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=559s)

![Recording frame at 574 seconds](https://ai.engineer/api/talk-essay-media/c78e885197f7f02566b05ee4c54f09206f32042ae297f817adde65e6f6bf68ce)

Recording frame at 574 seconds

This is a speaker-reported comparison of one sizable task; the recording does not supply enough task, harness, or evaluation detail to treat the numbers as a general benchmark. Its useful mechanism is narrower: if multiple sessions repeatedly search for the same architecture, ownership, and business rules, supplying that context up front can remove duplicated search work and leave more of the token budget for the requested change. [9:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=559s)

The reported outcome was roughly 50% fewer tokens, faster triage, and better answers because the model had business context. The token result is easiest to explain; the recording asserts the answer-quality improvement but does not provide an independent quality measure. [9:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=589s)

### Three open-source components expose the underlying techniques

[9:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=589s)

The first tool maps engineering relationships from GitHub activity. It identifies what a person commits, where they commit, and who reviews their work, then distills that activity into an experts graph. Team labels can optionally be added with an OpenAI or Anthropic API key, but the underlying relationship construction is deterministic. This graph gives the earlier auth request a concrete starting point: locate the requester, follow their collaboration and expertise links, and focus research on the parts of the organization where the request is likely to belong. [10:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=619s)

The second tool, the **repo rules agent**, discovers rules files across a repository, checks the extracted instructions and severities, and flags duplicates or other problems. The resulting index is queryable, so an agent can retrieve the rules relevant to its current work instead of loading every instruction file or searching several overlapping conventions independently. [10:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=649s)

The third tool demonstrates why semantic retrieval is only part of a context engine. Ask, “What are the open PRs that I worked on in the last week with authentication?” Similarity search alone cannot reliably apply the identity, status, time, and topic constraints. The workshop builds a relational query engine in six stacked pull requests: the agent discovers a schema, writes a query, and runs that query deterministically to retrieve structured records. RAG remains useful for finding semantically related documents; structured queries answer questions whose meaning depends on filters and relationships. [11:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=679s)

### Context becomes shared business infrastructure

[12:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=737s)

The same engine can support work outside code generation. Waselnuk describes customer-success teams using context to address tickets as they arrive and salespeople querying organizational information in the field. These workflows differ in their actions, but they need the same underlying abilities: retrieve current evidence, interpret it for the requester, and return it without violating permissions. [12:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=737s)

![Recording frame at 751 seconds](https://ai.engineer/api/talk-essay-media/48f476ee1ee33b39ce15ff6f1bdd16c50121d540f467312606dfa3b83ac8d5da)

Recording frame at 751 seconds

Unblocked also built a readiness tool that asks questions about a team’s current use of AI and maps it onto the earlier adoption curve, then suggests ways to move toward more scalable agent use. That closes the loop with the opening warning: adding parallel or background agents before giving them organizational context multiplies correction and review costs. [12:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=768s)

The closing claim is blunt: the important gap is no longer intelligence but context. Within this talk’s scope, that means surrounding capable models with current organizational knowledge, relationships, permissions, conflict handling, and economical delivery. Better models can improve implementation, but they do not automatically inherit the years of rejected PRs, meetings, incidents, and local judgment that made an experienced teammate effective. [13:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=791s)

## Key takeaways

- Agent errors become more expensive as teams move from autocomplete toward parallel and background execution, because misunderstandings survive longer before a human catches them.
- Static Markdown and MCP connections are useful, but maintenance, tool selection, and satisfaction-of-search failures keep them from supplying organizational understanding by themselves.
- A context engine must combine cross-system retrieval with identity-aware relevance, conflict resolution, permission enforcement, and token-efficient delivery.
- Semantic retrieval and structured querying solve different problems: RAG finds related material, while deterministic queries handle identities, relationships, statuses, and time ranges.
- The reported 21-million-versus-10.8-million-token comparison illustrates the potential cost of repeated discovery, but it is a single speaker-reported task rather than a general benchmark.

## Resources

- [Engineering Social Graph](https://github.com/unblocked/engineering-social-graph?utm_source=aie+worldfair) — Builds a developer collaboration and expertise graph from GitHub pull-request and Git history. Useful for grounding personalized retrieval in reviewers, teams, code areas, and recent working relationships.
- [Repo Rules Agent](https://github.com/unblocked/repo-rules-agent?utm_source=aie+worldfair) — Discovers coding-instruction files, extracts and indexes their rules, merges near-duplicates, flags potential conflicts, and supports task-scoped queries instead of injecting every rule into every prompt.
- [Beyond RAG: Structured Query Synthesis for Engineering Context](https://github.com/unblocked/document-query-engine?utm_source=aie+worldfair) — A six-step workshop for turning natural-language questions into validated MongoDB aggregation pipelines over GitHub PR and issue data, including schema discovery, identity resolution, validation, and retry.

## Related talks

- [Stop babysitting your agents: building a context engine for mergeable code](https://ai.engineer/talks/BiG2ssibKGc-stop-babysitting-your-agents-building-context) — Waselnuk develops the same context-engine argument with a detailed mergeability comparison, permission model, and MCP-assisted planning demonstration.
- [No Vibes Allowed: Solving Hard Problems in Complex Codebases](https://ai.engineer/talks/rmvDxxNubIg-context-engineering-for-complex-codebases) — Dex Horthy presents a complementary workflow for researching and planning changes in established codebases while deliberately controlling agent context.
- [Building agents is trivial now, context is the next frontier](https://ai.engineer/talks/HvMyYLTfvhg-building-agents-is-trivial-now-context-is) — Jeff Ng uses an incident-related example to show how unattended agents can miss organizational history even when connected to relevant systems.

## Slides

### Here’s how to fix “You’re absolutely right!”

Brandon Waselnuk introduces context engineering and plans to share open-source tools and techniques for addressing the familiar “You’re absolutely right!” response.

[0:20](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=20s) [Title slide for Brandon Waselnuk’s talk.](https://assets.ai.engineer/sites/aiecode2025/sha256/39adb1b64a4a35a0e2a08341089b7e0137fcefd70a68f054713c81c5c6be54bc.webp)

Here’s how to fix
“You’re absolutely right!”

Brandon Waselnuk

Unblocked

### AI-generated code should feel familiar

Waselnuk sets a goal for AI-generated code: it should feel as if a longtime teammate wrote it.

[0:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=48s) [AI-generated code should feel like it was written by someone who has been on your team for years.](https://assets.ai.engineer/sites/aiecode2025/sha256/afaed71719014c72a1b43cf5cff0685864493b231501658d23a5ce78b00206f8.webp)

AI-generated code should feel like it was written by someone who's been on your team for years.

Unblocked

### You have been the context engine

Developers build context over time by asking questions, shipping PRs, getting feedback, and attending meetings.

[1:20](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=80s) [For years, you have been the context engine.](https://assets.ai.engineer/sites/aiecode2025/sha256/94135d587688c8035512e1e7425ccb94b625c732fbd1886f728dcb6213ddd995.webp)

For years, you have been the context engine.

Unblocked

Interface text:
AIE WF26 — Your agents lack...
12:05 PM - 12:25 PM
Take notes

### Remember how you built context

To answer how auth validation works, people piece together scattered evidence: Slack says internal traffic skips the check, while Confluence says all requests check tokens. Experience adds context over time.

[1:32](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=92s) [Examples of outdated changes and documentation, unknown status, tribal knowledge, symptoms and conflicting information surround a question about authentication validation.](https://assets.ai.engineer/sites/aiecode2025/sha256/47ea273e0d37088738955ed44c42244fb81dd142fb83a75097c8a9e47052b886.webp)

Remember how you built context:

Outdated change
GitHub — 14d ago
Refactor auth flow — #3243
Replaced legacy JWT flow.

Unknown status
Jira — 5d ago
UNB-1337
Rejecting valid sessions.

Outdated docs
Notion — 6mo ago
System Architecture
Tokens validated by Auth Service.

Tribal knowledge
Slack — 3d ago
#eng-discussion
Internal traffic skips the check.

Symptom, not cause
Datadog — 4 hrs ago
auth.latency.p95
Spikes correlated with deploys.

Conflicting info
Confluence — 1mo ago
Auth-Service Overview
All requests check tokens.

How does auth validation work?

Unblocked

Interface text:
AIE WF26 — Your agents lack...
12:05 PM - 12:25 PM
Take notes

### Agents have the same context problem

Each new agent session starts without context on how the company operates. Waselnuk argues that getting that context wrong early compounds the cost as agents scale.

[1:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=108s) [A slide states that agents have the same problem and the cost compounds.](https://assets.ai.engineer/sites/aiecode2025/sha256/11fc475d04106cdddd2832455072b92a1766d99b49c6a236e0839c7c57a7df38.webp)

Now your agents have the same problem. Only the cost compounds.

Unblocked

Interface text:
AIE WF26 — Your agents lack...
12:05 PM - 12:25 PM
Take notes

### Your agents need a context engine

The maturity model moves from tab completion to agent teams as autonomy increases. Waselnuk argues that agents need a way to ask for context when humans are no longer checking each step.

[2:04](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=124s) [Eight rising bars show AI maturity from tab completion to agent teams, with increasing autonomy.](https://assets.ai.engineer/sites/aiecode2025/sha256/db02db42cbc6fb9fcdd12d17313490d0c83b862bbcd1e1c3dde27fe3966614a7.webp)

Your agents need a context engine
Levels of AI maturity in engineering organizations

AUTONOMY

1. Tab complete
2. Agent IDE
3. Context engineering
4. Compounding engineering
5. MCP + skills
6. Harness engineering
7. Background agents
8. Agent teams

YOU ARE THE CONTEXT
CURATED CONTEXT
CONTEXT LAYER

Unblocked
Stages adapted from Basssim Eledath

Interface text:
AIE WF26 — Your agents lack...
12:05 PM - 12:25 PM
Take notes

### The cost of bad context compounds

As agents gain autonomy, a bad completion can grow from a quick correction into repeated prompting, review work, or silent failure when nobody is in the loop.

[3:54](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=234s) [Four cards show escalating consequences: caught completions, repeated prompting, review work, and silent failure.](https://assets.ai.engineer/sites/aiecode2025/sha256/fb7b64cdea9a80c3dbfdb1d47b85334ff45215f5b8c415bf7eafa1289e66bf57.webp)

The cost of bad context compounds with every step toward autonomy

TAB COMPLETE
You catch it
One wrong completion. You hit escape, move on.

AGENT IDE
Doom loops
You re-prompt 3–5 times. Tokens burned, time lost.

PARALLEL AGENTS
Review tax
Five PRs land. Your senior engineers spend the day reviewing instead of building.

WITHOUT YOU IN THE LOOP
Silent failure
Wrong code ships. Wrong ticket enriched. Nobody saw it happen.

Unblocked

### Common approaches that don’t work

Waselnuk says shared context repositories go stale, while agents may overlook MCP tools or stop searching after the first plausible result.

[4:09](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=249s) [A slide describes stale shared rules and the MCP plateau as common context problems.](https://assets.ai.engineer/sites/aiecode2025/sha256/9a23d645280065c7a96fb57c9dc9d730fff7f4f5cdab563c9a7f6f55557f30bd.webp)

Common approaches that don’t work

• Curated-context trap
The shared rules repo rots faster than you can update it.

• MCP plateau
More servers, same output quality — agents don’t know when or why to call them.

Unblocked

### The problem: access ≠ understanding

An agent may proceed after finding an architecture record and miss a more recent Slack decision. Access to information alone does not ensure understanding.

[5:41](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=341s) [Slide stating that access does not equal understanding.](https://assets.ai.engineer/sites/aiecode2025/sha256/6f75aacd181bfe3f01d9fd119cbb8e72456cd34ce3a1ab6554bdb8e8eb9b4361.webp)

The problem: access ≠ understanding

Unblocked

### What your agent can’t see

The iceberg puts code that compiles above the waterline and team knowledge below it. Waselnuk warns that missing a rollout procedure, such as turning off a feature flag, can lead to a production outage.

[5:50](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=350s) [Iceberg diagram contrasts code that compiles with hidden context such as original intent, past decisions and incident learnings.](https://assets.ai.engineer/sites/aiecode2025/sha256/e8018e6d327e2107d90977d85e938f3705be42c6fa518590a48208545ef2a85f.webp)

CODE THAT COMPILES

ORIGINAL INTENT
TEAM CONVENTIONS
PAST DECISIONS
WHY WAS IT BUILT THIS WAY
MIGRATION PLANS
TESTING STANDARDS
REJECTED APPROACHES
SLACK DECISIONS
ARCHITECTURE RATIONALE
DESIGN DOC WITH TRADE OFFS
INCIDENT LEARNINGS

What your agent can’t see

imgflip.com
Unblocked

### Why your team needs a context engine

A context engine should use your role to find relevant information, resolve conflicts across sources, respect permissions, and deliver context to the model when needed.

[6:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=371s) [Four bullets describe relevance, conflict resolution, permissions and timely delivery of context.](https://assets.ai.engineer/sites/aiecode2025/sha256/3b6bf612a70f5a6000af0cc7a913270ebc7a2a70a38b26205cb3d4bff04bedb7.webp)

Why your team needs a context engine

• Understands who you are and which information actually matters.
• Resolves conflicts between different sources (code, docs, Slack, tickets).
• Respects permissions and governance.
• Delivers the right context to the right model at the right time.

Unblocked

### Six requirements for a real context engine

Waselnuk’s context engine combines signals across sources, resolves conflicts, and retrieves information suited to the user and task. It also compresses context and enforces permissions across systems.

[7:55](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=475s) [Table lists unified context, targeted retrieval, conflict resolution, personalized relevance, token optimization and permission enforcement.](https://assets.ai.engineer/sites/aiecode2025/sha256/7e61f31c3051ada62aa64dd70fb70191476eadf1023047e31c1fdb011a996c16.webp)

Six requirements for a real context engine

Unified system context
Merges signals from all sources before delivery to the agent.

Targeted retrieval
Retrieve only the information agents need from the graph.

Conflict resolution
Recency + authority signals resolve contradictions automatically.

Personalized relevance
Scoped to your repos, teammates, and work history.

Token optimization
Ranked, compressed context — no waste in the prompt window.

Permission enforcement
Permissions and policies enforced automatically across systems.

Unblocked

### Open source software

Brandon Waselnuk asks the audience to get their cameras ready for three QR codes.

[9:58](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=598s) [Open source software section slide with a camera illustration.](https://assets.ai.engineer/sites/aiecode2025/sha256/1477538be8d0907e9ce545899b2af083ee5582807747e45770b1369f7d1a49fa.webp)

Open Source Software
(GET YOUR CAMERAS READY)

Unblocked

## Full transcript

[0:01](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=1s) [music]

[0:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=12s) >> Good afternoon.

[0:13](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=13s) I hope you're all having a lovely day

[0:15](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=15s) here at AIE.

[0:16](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=16s) We've had great weather, though the UV

[0:18](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=18s) has been like nine. So, hopefully you

[0:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=19s) put your sunscreen on your being

[0:21](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=21s) appropriate adults.

[0:22](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=22s) I'm here to talk to you about context

[0:24](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=24s) engineering, and I have the good fortune

[0:25](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=25s) of following AJ from LinkedIn because he

[0:27](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=27s) talked a lot about the system that we

[0:28](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=28s) actually design and sell to other

[0:29](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=29s) solutions. And I'm going to give you a

[0:31](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=31s) bunch of open source tools. So, if you

[0:33](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=33s) watch that last talk just before me,

[0:35](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=35s) you're going to get a bunch of tool

[0:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=36s) chance you can go mess around yourself,

[0:37](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=37s) and I'll teach you a bunch of techniques

[0:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=38s) today. The goal, of course, is to fix

[0:41](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=41s) your absolutely right.

[0:43](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=43s) I think they've taken that out of the

[0:44](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=44s) prompts now, so it just says you're

[0:45](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=45s) right or other things, but I'm sure

[0:47](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=47s) you've all been there.

[0:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=49s) So, I'm Brandon.

[0:50](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=50s) I work at Unblocked. Uh yes, I have a

[0:52](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=52s) coconut. We've been giving these away

[0:53](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=53s) for fresh context, fresh fresh coconuts.

[0:57](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=57s) But, the thing that I want to talk to

[0:58](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=58s) you about is with these models,

[1:00](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=60s) especially with Meth O'Clock models, I

[1:01](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=61s) think Fable 5's coming back today, so

[1:03](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=63s) they say.

[1:04](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=64s) You can watch my Grain Call recording

[1:06](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=66s) try to book this.

[1:07](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=67s) We'll ignore it.

[1:09](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=69s) But, what I want you to do is to think

[1:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=71s) about the fact that with these tools,

[1:13](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=73s) AI-generated code should feel like it

[1:15](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=75s) was written by someone who's been on

[1:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=77s) your team for years.

[1:21](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=81s) So, to get in the right headspace, for

[1:23](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=83s) years you have to consider that you have

[1:25](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=85s) been the context engine.

[1:27](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=87s) How did you do that?

[1:29](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=89s) You built context by going to work

[1:32](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=92s) and asking questions,

[1:34](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=94s) shipping PRs and getting them rejected,

[1:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=96s) going to meetings, and all this slowly

[1:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=98s) over time built up the engine that is

[1:39](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=99s) your brain.

[1:40](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=100s) You understand how it works here. You

[1:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=102s) know how stuff gets shipped. You were on

[1:44](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=104s) call that night when you took prod down

[1:46](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=106s) and why that happened.

[1:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=108s) The problem is

[1:50](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=110s) that these agents have this exact same

[1:51](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=111s) problem. Every time you create a new

[1:53](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=113s) terminal session with an agent in it,

[1:55](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=115s) it's very intelligent, but it doesn't

[1:57](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=117s) have any context on how your company

[1:58](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=118s) operates. So, it needs to get that

[2:00](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=120s) somehow.

[2:02](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=122s) The problem is as you move these agents

[2:03](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=123s) up in scale,

[2:05](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=125s) that cost compounds if you get it

[2:07](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=127s) incorrect at the beginning. The leverage

[2:10](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=130s) of context and content

[2:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=132s) We're just going to fix this cuz I think

[2:13](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=133s) people want to take some photos.

[2:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=137s) Perfect.

[2:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=139s) That context issue will compound. So, at

[2:22](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=142s) the far left, we all remember the

[2:24](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=144s) age-old time of 2 years ago where we had

[2:26](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=146s) tab complete models that were pretty

[2:27](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=147s) cool. What happened is it popped up and

[2:29](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=149s) said, "Hey, do you want to tab this?"

[2:31](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=151s) And quickly in your head with your

[2:32](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=152s) context engine, you go, "No, that's

[2:33](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=153s) bad." Or you went, "Oh, sweet." You hit

[2:35](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=155s) tab. Nice.

[2:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=156s) As we move along the agentic kind of

[2:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=158s) adoption curve, what happens is you are

[2:41](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=161s) moving into more situations in which you

[2:43](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=163s) have agents running without a human in

[2:45](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=165s) the loop, or at least you wish you

[2:46](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=166s) didn't have to be in the loop.

[2:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=168s) What they need is some way to be able to

[2:50](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=170s) ask the questions they need when they

[2:52](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=172s) hit walls in order to write code or

[2:54](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=174s) solve or basically fix the issue and

[2:56](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=176s) ultimately output code that's mergeable

[2:58](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=178s) into your code base, especially with

[3:00](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=180s) many people here who actually work in

[3:02](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=182s) brownfield code bases that have been

[3:03](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=183s) around for a long time that run real

[3:05](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=185s) revenue across them, not just greenfield

[3:08](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=188s) fun projects.

[3:10](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=190s) So, that cost of bad context compounding

[3:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=192s) at the beginning is cheap. If you think

[3:14](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=194s) like shift left, finding a defect or a

[3:16](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=196s) bug, you want to find it as early as

[3:18](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=198s) possible. It's the same with context.

[3:21](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=201s) Cuz as you move across, you get into

[3:23](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=203s) doom loops. You usually ask your to do

[3:25](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=205s) something. It's like, "Hey, I did it."

[3:26](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=206s) And you're like, "No, man." And then you

[3:28](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=208s) correct and correct and correct. That's

[3:30](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=210s) wasted search tokens. It's also wasted

[3:32](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=212s) rework time.

[3:34](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=214s) And that is not acceptable with the

[3:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=216s) tokenomics we have coming.

[3:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=218s) And then as you move into parallel

[3:39](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=219s) agents, etc., you start hitting a review

[3:41](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=221s) tax. So, these AI code reviewers we're

[3:43](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=223s) trying to use, but again, key context is

[3:45](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=225s) important there so that those code

[3:47](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=227s) reviews are able to basically

[3:50](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=230s) understand how the operations of the

[3:51](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=231s) business are so it knows the business

[3:53](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=233s) logic and more.

[3:54](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=234s) And then finally, if your hope is to

[3:56](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=236s) move all the way out of the loop, you're

[3:58](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=238s) like background agents, get it done,

[4:00](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=240s) make no mistakes, you really need to

[4:02](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=242s) make sure that you have a context engine

[4:03](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=243s) so those agents can query it and get all

[4:05](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=245s) the answers they need so they can keep

[4:07](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=247s) operating in an effective way.

[4:10](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=250s) There are some common approaches that

[4:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=252s) don't work. They're basically like a

[4:14](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=254s) local maxima.

[4:15](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=255s) Two of the ones we see the most with our

[4:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=257s) hundreds of enterprise clients and

[4:18](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=258s) mid-market size businesses

[4:20](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=260s) is the curated context trap. If you've

[4:23](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=263s) ever sat down and taken a virtual file

[4:24](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=264s) system or maybe a local file system, you

[4:27](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=267s) put some markdown files in it and you're

[4:28](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=268s) like, here's all the context of this

[4:29](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=269s) project, it's how it works. You then

[4:31](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=271s) allow your agent to grep over that and

[4:33](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=273s) it gets a bunch of good data and then it

[4:34](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=274s) will perform better.

[4:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=276s) The issue is first, now you have to

[4:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=278s) distribute that so maybe you throw it up

[4:40](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=280s) in a GitHub and your team can grab it.

[4:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=282s) But then the next is that repo is going

[4:44](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=284s) to rot just like all the other docs you

[4:46](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=286s) wrote down and then who at your org is

[4:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=288s) the omnipotent one who has the taste to

[4:51](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=291s) curate this file or repo for literally

[4:53](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=293s) everyone in the org. So you start to hit

[4:55](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=295s) these issues.

[4:57](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=297s) The next is the MCP plateau.

[4:59](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=299s) This one is pretty clear. We have MCPs,

[5:02](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=302s) they're great. You can give it to your

[5:03](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=303s) agent and now it can basically get

[5:05](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=305s) information from another source system.

[5:07](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=307s) The problem is, of course, based on how

[5:10](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=310s) you write the server description, the

[5:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=312s) tool descriptions, your agent may never

[5:14](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=314s) call it even though it should have. Or

[5:16](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=316s) if it does, there's a known bias called

[5:18](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=318s) the satisfaction of search bias. What

[5:20](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=320s) that means is the agent, when it finds

[5:22](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=322s) the first piece of information that it

[5:24](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=324s) thinks is correct, it goes, "Oh, I have

[5:26](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=326s) what I need." and it proceeds.

[5:28](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=328s) In most organizations, there's a Slack

[5:30](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=330s) conversation from last night that says

[5:32](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=332s) you should be doing A instead of doing B

[5:34](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=334s) and the agent will never find it if it

[5:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=336s) found some architecture record first.

[5:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=338s) So it doesn't actually consider all of

[5:40](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=340s) the context.

[5:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=342s) The problem here is access to

[5:44](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=344s) information is not understanding.

[5:47](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=347s) So to deliver understanding to a model,

[5:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=349s) you have to do other techniques.

[5:52](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=352s) What I'm basically trying to say is

[5:54](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=354s) what your agent can't see is everything

[5:55](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=355s) below the waterline. It can 100% get

[5:58](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=358s) code that compiles, but that code that

[6:00](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=360s) compiles is taking down prod and you

[6:02](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=362s) have a P0 at 1:00 in the morning.

[6:04](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=364s) Because it missed the fact that you have

[6:06](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=366s) a certain rollout procedure, you're

[6:07](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=367s) supposed to turn off a feature flag,

[6:09](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=369s) whatever it might be.

[6:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=372s) So, your team needs a context engine

[6:14](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=374s) because what it should do is understand

[6:16](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=376s) who you are and where you work in an

[6:18](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=378s) organization. So, if I say to you, I

[6:20](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=380s) want to get off stood up, it knows where

[6:22](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=382s) I work, it knows where my get commits

[6:24](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=384s) are, it knows who reviews those commits,

[6:27](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=387s) and it understands that my context, it

[6:29](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=389s) can focus me, and then use that as a

[6:31](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=391s) trigger point to find the rest of the

[6:32](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=392s) information.

[6:34](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=394s) It resolves conflicts, as mentioned, an

[6:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=396s) old architecture diagram and last

[6:37](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=397s) night's Slack convo with the CTO,

[6:40](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=400s) which one is right? You need to use a

[6:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=402s) bunch of techniques to discern determine

[6:44](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=404s) that.

[6:45](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=405s) Respects permissions and governance, of

[6:46](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=406s) course. MCP allows us to use OAuth and

[6:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=409s) other scopes and SSO, but if someone

[6:51](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=411s) asks a question over here who's not

[6:52](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=412s) supposed to know about secret project A,

[6:56](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=416s) you need to make sure that doesn't leak

[6:57](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=417s) into the response.

[6:59](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=419s) And then finally, deliver the right

[7:00](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=420s) context at the right time to the model

[7:02](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=422s) in a token optimized way.

[7:04](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=424s) We have multiple surface areas because

[7:05](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=425s) human engineers still talk to Unblocked

[7:07](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=427s) all the time to get information they

[7:08](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=428s) need in Slack or otherwise, but then you

[7:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=431s) want token optimized responses if you're

[7:13](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=433s) just speaking machine to machine in

[7:14](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=434s) order to not waste a bunch of bold

[7:16](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=436s) classes on your token spend.

[7:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=439s) This is how an engine works. I'm going

[7:21](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=441s) to be brief on this, but basically on

[7:23](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=443s) the left-hand side,

[7:25](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=445s) you see all the data sources that are

[7:26](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=446s) coming in.

[7:27](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=447s) For us, we focus on engineering teams

[7:29](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=449s) and that's who uses us, as well as the

[7:31](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=451s) technically light teams around it, like

[7:33](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=453s) support, sales, and otherwise.

[7:35](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=455s) You ingest all that data, you get

[7:37](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=457s) real-time data from tools like your

[7:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=458s) instant management tool chain.

[7:40](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=460s) It comes into the engine, where that

[7:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=462s) engine is, it thinks at the bottom. I'll

[7:44](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=464s) expand on that slide in a moment. But

[7:46](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=466s) basically it uses these six key

[7:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=468s) characteristics. And then on the right,

[7:50](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=470s) you output the context to the exact

[7:52](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=472s) workflow in the manner that it is

[7:53](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=473s) needed.

[7:56](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=476s) Those six key points, as mentioned,

[7:58](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=478s) unified system context, you have to go

[7:59](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=479s) across the whole thing. At large orgs,

[8:02](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=482s) companies like LinkedIn scale, Workday,

[8:05](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=485s) General Motors, whatever, they need this

[8:07](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=487s) type of data. They need to understand

[8:09](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=489s) everything that's happening. And Threek

[8:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=491s) this morning actually talking about

[8:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=492s) Fable coming out potentially later

[8:14](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=494s) today,

[8:15](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=495s) he mentioned that you need to actually

[8:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=497s) provide a map and then let Fable

[8:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=499s) discover the territory. The way to help

[8:21](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=501s) confine that is making sure that these

[8:24](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=504s) models have access to all of the

[8:25](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=505s) context, because they will find your

[8:28](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=508s) unknown unknowns.

[8:29](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=509s) There are definitely things going on in

[8:31](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=511s) your company that you're just unaware

[8:32](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=512s) of, but would be really helpful for the

[8:34](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=514s) task you're trying to do.

[8:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=516s) That will move faster, but the targeted

[8:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=518s) retrieval, you should be able to if you

[8:39](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=519s) provide a link quickly, unfurl it, get

[8:41](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=521s) that document back and move along. So,

[8:44](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=524s) two tasks, deep research, go long,

[8:46](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=526s) that's fine, but you also need speed

[8:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=528s) when speed is required. Conflict

[8:50](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=530s) resolution, we already talked about

[8:51](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=531s) that, but one thing says do A, one thing

[8:53](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=533s) says do B, who is right?

[8:55](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=535s) Personalized relevance, who am I, where

[8:57](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=537s) do I work, what am I working on?

[8:59](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=539s) That token optimization, making sure the

[9:01](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=541s) response is good and effective and

[9:03](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=543s) doesn't bloat the window.

[9:05](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=545s) And then permission enforcement, of

[9:06](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=546s) course, OAuth, you shouldn't see it, you

[9:08](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=548s) shouldn't see it.

[9:10](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=550s) What we did with some tests is we

[9:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=551s) actually ran the exact same prompt to

[9:13](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=553s) the same model and one with context and

[9:16](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=556s) one without. This is the wall clock time

[9:18](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=558s) savings.

[9:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=559s) And then 2 hours, which is great. And

[9:22](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=562s) then the tokens savings. So, it was a

[9:24](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=564s) sizable task, it took about 21 million

[9:27](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=567s) tokens without and then 18, or sorry,

[9:29](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=569s) 10.8 million tokens with it.

[9:32](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=572s) This is the type of experience that you

[9:33](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=573s) typically see when you're using a

[9:35](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=575s) context engine, cuz the majority of

[9:37](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=577s) those wasted search tokens where it has

[9:39](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=579s) to grab at the beginning of every

[9:40](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=580s) session to understand and discover

[9:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=582s) things are no longer there when it's

[9:43](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=583s) hydrated with context. Hydrated.

[9:47](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=587s) And then, as you move forward, you get

[9:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=589s) these types of outcomes.

[9:51](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=591s) 50% fewer tokens, faster triage, and the

[9:53](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=593s) answer quality is actually better

[9:55](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=595s) because it knew what was going on inside

[9:57](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=597s) of the business.

[9:59](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=599s) Now, this next part,

[10:01](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=601s) you'll probably want to photo. If you

[10:02](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=602s) don't know, you can actually take a

[10:03](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=603s) picture of a QR code and then later in

[10:05](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=605s) photos tap on it and then load the link

[10:07](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=607s) so you don't need to float here cuz I'm

[10:08](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=608s) going to give you three QR codes.

[10:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=611s) This first one is for the social comment

[10:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=612s) network. I'll pop that up so you can

[10:13](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=613s) take a photo.

[10:15](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=615s) But, this is an open source tool that

[10:16](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=616s) we've got that actually, using all

[10:18](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=618s) deterministic programming, goes over

[10:20](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=620s) your GitHub and understands who works on

[10:22](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=622s) your team. This is my real team. We

[10:23](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=623s) called Rasheem the machine cuz he ships

[10:26](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=626s) like crazy. But, on the right, you can

[10:28](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=628s) see who he commits, where he commits,

[10:29](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=629s) who's reviewing his work. And then in

[10:31](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=631s) those tabs, you can find a distilled

[10:33](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=633s) experts graph. You get full coverage of

[10:35](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=635s) what's going on in your business. And if

[10:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=636s) you optionally add one of the API keys

[10:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=638s) for either OpenAI or um Anthropic, it'll

[10:41](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=641s) um determine what your teams are by

[10:43](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=643s) doing some labeling for you.

[10:45](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=645s) It's a really cool tool to understand

[10:46](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=646s) where your team works and get that

[10:47](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=647s) social network in there in order to

[10:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=649s) focus the context engine if you're going

[10:50](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=650s) to be building these tools yourself.

[10:53](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=653s) The next is called the repo rules agent.

[10:56](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=656s) This is a sample from our real code

[10:58](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=658s) base. I'm going to pop that up anyway so

[10:59](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=659s) you don't need to talk to the thing, but

[11:01](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=661s) in short, what it does is discover all

[11:04](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=664s) the places your team has written rules

[11:05](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=665s) files, checks them all, and then tells

[11:08](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=668s) you what severities you've given,

[11:10](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=670s) what other things you've given. Should I

[11:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=671s) just switch to this?

[11:13](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=673s) It tells you what it Whoa, hey.

[11:15](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=675s) It's good to meet you all.

[11:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=677s) Basically, it will find all the rules

[11:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=679s) that are inside of your repo and then

[11:21](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=681s) tell you if you have duplicate issues or

[11:23](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=683s) others problems and then you can grab

[11:25](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=685s) over it as an index. So, that index can

[11:27](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=687s) be called and you can dedupe and it'll

[11:28](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=688s) help improve um your retrieval of

[11:30](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=690s) context.

[11:32](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=692s) And then finally, on Monday we delivered

[11:34](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=694s) this workshop, which was going beyond

[11:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=696s) rag and taught how to build a relational

[11:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=698s) context engine from scratch.

[11:41](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=701s) So, if you scan that, you'll get the

[11:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=702s) full workbook. It has six PRs stacked

[11:44](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=704s) that teach you how to walk through doing

[11:46](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=706s) this. But in short, rag is an incredible

[11:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=708s) technique and you want that. But the

[11:50](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=710s) other half of the problem is what people

[11:52](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=712s) actually ask is, "What are the open PRs

[11:55](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=715s) that I worked on in the last week with

[11:57](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=717s) authentication?"

[11:58](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=718s) Rag cannot answer that question alone.

[12:01](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=721s) You need queries. So, this shows you how

[12:03](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=723s) to do a

[12:04](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=724s) schema-less basically look up that

[12:06](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=726s) allows the agent to discover a schema

[12:09](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=729s) and then write queries against it

[12:10](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=730s) deterministically in order to get that

[12:12](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=732s) type of relational data out.

[12:14](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=734s) Very useful technique.

[12:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=737s) Use cases of a context engine, of

[12:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=739s) course, do go beyond code generation.

[12:21](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=741s) This is, you know, where we live a lot,

[12:23](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=743s) a lot of our customers spend their time.

[12:25](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=745s) But it's amazing to see what happens

[12:26](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=746s) when a bunch of other people around the

[12:28](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=748s) business start picking up these tools,

[12:31](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=751s) customer success people solving tickets

[12:33](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=753s) right at the time that it comes in from

[12:34](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=754s) a customer.

[12:36](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=756s) We've got sales people closing deals

[12:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=758s) earlier in their quarter because they're

[12:39](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=759s) able to just query the Unblocked context

[12:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=762s) engine on the fly while in the field.

[12:44](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=764s) And so many more.

[12:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=768s) What you can also do is if you saw that

[12:49](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=769s) curve chart earlier where I talked about

[12:51](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=771s) the levels, we've built a fun little

[12:52](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=772s) tool where basically an LLM will quiz

[12:54](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=774s) you and ask you about what's going on

[12:56](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=776s) and then it will map you to exactly

[12:57](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=777s) where you are and then tell you some

[12:59](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=779s) techniques about how to level up through

[13:01](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=781s) that if you are looking to basically

[13:03](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=783s) compound your capabilities and ship with

[13:05](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=785s) AI tools at scale. It's

[13:07](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=787s) readiness.unblocked.com.

[13:11](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=791s) The gap is not intelligence any longer.

[13:13](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=793s) It's context. We will continue to get

[13:16](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=796s) incredible models like Mythos as it's

[13:17](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=797s) been grown by Anthropic and I'm sure

[13:19](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=799s) Soul once I'm allowed to see it. I will

[13:22](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=802s) get it. Happy Canada Day, by the way.

[13:24](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=804s) But what's happening is it's about the

[13:26](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=806s) context you surround these models with

[13:28](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=808s) in order for them to be effective and

[13:31](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=811s) token efficient inside of your

[13:32](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=812s) organization.

[13:35](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=815s) So, I have a question slide, but I'm not

[13:38](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=818s) sure I'm allowed.

[13:40](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=820s) Nope. So, what you'll do is come meet me

[13:42](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=822s) at booth P16. You can look for the

[13:45](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=825s) coconut.

[13:46](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=826s) It'll be great to hang out with all of

[13:47](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=827s) you and get into details here if you

[13:48](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=828s) need it. Thank you for your time.

[13:51](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=831s) >> [applause]

[14:06](https://www.youtube.com/watch?v=KcVkq5L-0f0&t=846s) [music]
