AI Engineer Summit 2023
The Age of the Agent
Loading slides…
Read the talk
The Age of the Agent
From a calendar correction to agents building integrations, Flo Crivello explores how delegating actions could change who can build a business—and what still goes wrong.
From a talk by Flo Crivello
What if one person could have Coca-Cola’s impact?
Could one person have as much business impact as Coca-Cola? Flo Crivello opens with that possibility, placing it five to ten years into the future he imagines. The precedent is media: reducing the cost of reaching an audience changed who could participate and what they could create.
Consider the institutional machinery behind Oprah’s media empire: executive pitches, fundraising, a production crew and camera operators. YouTube made starting possible with a laptop. Crivello describes MrBeast as having greater reach than Oprah or the Super Bowl. The slide illustrates the scale with 10M viewers for Oprah and 189M subscribers for MrBeast, although viewers and subscribers are different measures, not a like-for-like audience comparison.
MrBeast began with a laptop; Ryan’s World makes the change in formats even more striking. Crivello describes Ryan starting toy unboxing and review videos, and having $100 million in wealth at the time of the talk. Those are his biographical and wealth claims, rather than a documented accounting of earnings or brand revenue. The larger point is that reviewing toys became a media business that the old gatekeeping system was unlikely to commission.
Removing gatekeepers changes the nature of the output, not just its price. Talk shows and sports broadcasts fit familiar categories. MrBeast’s productions and Ryan’s toy videos show how unfamiliar formats can emerge when starting becomes easier. Applied to business, the prediction is that reducing operational friction will let stranger, more creative ideas reach the world.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A business needs actions, not just generated content
Before Lindy, which he calls his second startup, Crivello founded Teamflow. He initially thought starting a company meant building a compelling product and bringing it to market. Then came the prerequisites: lawyers and incorporation, bankers and an account, investors and fundraising, recruiters and hiring. Operating the business did not make that overhead disappear.
Generative tools for copywriting and images were useful, but they left his underlying problem intact: starting a business was still painfully laborious. His distinction is between producing content and performing the work that moves a business forward. Agentic AI attracted him because it could take actions and automate the menial parts of that work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Start with the work nobody wants
The image of administrative drudgery is Milton in Office Space: basement work, TPS reports and, eventually, burning the building down. Crivello uses the ending figuratively as a rejection of the status quo. His reversal of the usual automation anxiety is that people have been doing robots’ jobs—work that few would choose as the purpose of their working lives.
Crivello cites fifteen administrative hours per week for the average US manager and $459 billion in annual US costs, which he compares with Norway’s GDP. The historical basis matters: the matching 2015 ServiceNow research surveyed managers at large companies, rather than all US managers, and its published report extrapolated approximately $575 billion in administrative wage costs. That does not establish recoverable savings or substantiate the talk’s $459 billion figure and GDP comparison.
Lindy’s starting point is consequently narrower than running an entire business: an AI employee acting as a personal assistant. Its first targets are three recurring sources of overhead—calendar, email and meetings.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A calendar correction becomes a preference
The calendar demonstration uses actual product screenshots. A request to find half an hour every week with Eric produces a meeting named for both participants. The interesting behavior is not merely finding a slot: the title reflects feedback from an earlier interaction.
In that earlier exchange, Crivello asked for thirty minutes with Eric the next day. Lindy created an invite titled Meeting with Flo, which was unhelpful when so many of his meetings had that title. He corrected it to include both names. Lindy renamed the meeting and saved the preference for future scheduling; the screenshot shows the resulting title as Flo : Eric. The correction changes both the current artifact and future behavior.
Crivello extends this example into a broader claim: users can supply arbitrary preferences, even complex ones, and Lindy will remember and honor them. Scheduling can also begin outside the assistant’s own interface by CC’ing Lindy on an email. The demonstrated preference update establishes the interaction pattern; the talk does not expose how that memory is stored internally.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Draft replies and prepare the next meeting
Email assistance adds recipient-specific behavior. Lindy prepares replies in the user’s voice, accounting for the difference between writing to a partner and writing to investors. Crivello describes opening Gmail each morning to find drafts ready for review. The delivery boundary here is a draft awaiting the user, not an automatically sent message.
Meeting preparation combines timing with information retrieval. His instruction requests three items five minutes before each meeting:
Five minutes before every meeting, send me:
- The Zoom link
- The LinkedIn profiles of the people I’m meeting with
- A summary of my last few emails with them
This turns a repeated preparation routine into a standing request with a trigger, several information sources and a deliverable.
Crivello says the team did not individually build these behaviors as dedicated features. Instead, it built a framework that lets an AI pursue arbitrary goals using arbitrary tools. On that account, the assistant’s useful behavior comes from combining an instruction with available capabilities, rather than requiring a separately programmed feature for every request.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A working routine can deliver false information
Successful demonstrations can hide an agent’s most consequential failures. After criticizing cherry-picked AI presentations, Crivello describes asking Lindy to email him an interesting new vocabulary word every morning. The routine ran, the emails arrived and he began using the words.
Then came pulchrivity, presented as a term for meandering through a conversation without a fixed direction. Suspicious, he searched Google and found that it did not exist. He then checked the earlier words and reports that none of them existed either. His joke about Lindy poisoning his brain captures the failure precisely: delivery had succeeded repeatedly while the content was fabricated, and he had already acted on it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
One assistant is only the beginning
When the system works, the envisioned computing experience becomes conversational. Instead of spending the day manipulating a computer to complete administrative tasks, the user describes what should happen and concentrates on the work they do best. The desired result is sustained focus while the unwanted coordination happens automatically.
That still does not reach the opening ambition. A personal assistant can remove overhead without giving one person the productive capacity of a major company. The next step is a society of agents: multiple Lindys working together on the user’s business goals.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A toaster, an iOS app and a rocket
Even simple manufactured objects embody collective work. Crivello invokes Thomas Thwaites’s Toaster Project, an attempt to make a toaster from scratch, to expose the difference between one person’s capabilities and the capabilities of an industrial system.
In Crivello’s retelling, the handmade toaster took six months and about $2,000, or roughly $50,000 with his estimated value of the maker’s time, versus $25 for a commercial toaster. Thwaites’s own portfolio records different figures: nine months and £1,187.54, compared with a £3.94 Argos toaster. The useful contrast survives those discrepancies: a costly, rough handmade result beside an inexpensive ordinary appliance produced through specialization and shared infrastructure.
The analogous model request is to ask GPT-4 to design an entire iOS app, implement it and publish it to the App Store. If it fails, one possible response is to wait for GPT-5, GPT-6 or GPT-7. These names appear in the talk as anticipated successors, not as a comparison of their measured capabilities.
But an individual person’s inability to make a rocket does not prove that humans cannot make rockets. Crivello applies the same distinction to models: an individual agent’s limit need not be the coordinated system’s limit. Lindy’s multi-agent framework is presented as a way to pursue that possibility by letting agents work together toward a shared goal. The toaster and rocket comparisons motivate the architecture; they do not establish how much capability a particular agent team gains.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Agents building integrations for other agents
The concrete example is Lindy helping build itself. An assistant needs integrations with services such as Slack, Twilio and Google Sheets to act in a user’s working environment. Building those integrations creates a recurring engineering workload that the team is trying to delegate to a society of Lindys.
The workflow starts with an instruction such as Build a Slack integration and divides the work into specialized responsibilities:
- Tool creation: A top-level Lindy receives the integration request.
- Documentation research: Another Lindy finds the OpenAPI specification and online documentation.
- Work allocation: An engineering manager divides the task among engineering agents.
- Implementation: Software engineers work on the pieces, with a dedicated engineer for authentication because it has particular implementation pitfalls.
- Quality assurance: A QA engineering agent receives the resulting work.
The division is more specific than asking several agents the same question: each role contributes a different part of the integration.
QA provides the feedback loop. If the implementation does not work, the QA agent sends it back to the software engineer. If it works, the agent submits a pull request. The workflow ends at PR submission; the talk does not describe an automatic merge or deployment.
Crivello estimates that the integration-building system is 70–80% of the way there. That is his assessment of prototype progress, not a measured task success rate. The example therefore shows the intended organization of research, implementation and verification without presenting the entire system as complete.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The remaining skill is directing the work
The coordinated system brings the talk back to one person in a San Francisco studio with Coca-Cola-scale business impact. Crivello’s equalizer is access to expertise: where an exceptional chief marketing officer might now work for Apple, Nike or Coca-Cola, he predicts that the best future CMO will be an AI—and makes the same prediction for designers and engineers. Those specialists would work for individual users, not only organizations able to recruit them.
His metaphor is a lever of infinite strength available to everyone. In that forecast, the decisive human skill becomes knowing how to use it: directing available capability toward a chosen result. The closing invitation is to imagine building without today’s constraints of time, money, team or network—just a person, a laptop and their Lindys.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Thomas Thwaites explains his attempt to make a toaster from raw materials and the industrial infrastructure it exposed.
Further reading
Crivello describes teams of Lindies, including agents that research API documentation, divide work, write code and perform QA.
- Slack API specificationsRepository
Slack’s archived OpenAPI and AsyncAPI specifications illustrate machine-readable inputs for integration development.
Updates since the talk
Current first-party documentation for building and using Lindy agents.
Full transcript
Loading transcript…