AI Engineer Summit 2023
The Age of the Agent
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 a 25-year-old 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 at three, and being twelve with $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 a child 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 the 25-year-old 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.
Read the complete timestamped transcript
- 0:00
[upbeat music] I'm going to talk about the future that awaits us, and not a super distant future either.
- 0:21
Like, I'm talking about five, 10 years, certainly within our lifetimes.
- 0:26
The future that awaits us once agents have fully realized their potential.
- 0:33
If I had to describe it in one sentence, I'd say that it's a world where a [REDACTED:age] can have the same or more business impact as The Coca-Cola Company.
- 0:46
It sounds insane when you say it this way, but there's actually a precedent to it. It's happened before with media.
- 0:53
Consider what Oprah had to do to build her media empire, right? She had to go and pitch a bunch of, a bunch of CNN executives in some stuffy room and raise money, hire her crew, find cameramen.
- 1:07
And obviously, the internet and apps like YouTube have brought that friction down to zero.
- 1:15
And it's brought that friction down so much that it actually sounds like a joke. Consider the top YouTuber, MrBeast, right? [laughs]
- 1:23
He's got a much greater reach than Oprah. He, he actually has more of a reach than the Super Bowl,
- 1:31
and it's just him and his laptop is how he got started. So it happened before. Even weirder, Ryan's World.
- 1:39
He started when he was [REDACTED:age], making videos on YouTube of him un- unboxing and reviewing toys.
- 1:47
Today, he's [REDACTED:age], and his wealth's a hundred million dollars.
- 1:54
My point is that once you bring the friction down to zero and once you remove the gatekeepers,
- 2:01
you don't just get the same kind of content except cheaper.
- 2:06
The nature of the content changes when you remove the gatekeepers.
- 2:12
Right? Look at Oprah, look at the Super Bowl. That makes sense, right? It's like a talk show. It's like a sports game.
- 2:18
Ryan's World and MrBeast are just weird. And so my point is that we're about to see the exact same transformation happen to the world of business at large.
- 2:29
We're gonna take the friction down to zero, and as a result, we are going to see much weirder, more creative ideas come to life.
- 2:41
You know, Lindy is my second startup. Before it, I had another one called Teamflow. And, um, I remember when I started it, I had a perhaps naive understanding of what starting a business entailed.
- 2:50
I thought it was all about building a cool product and bringing it to market.
- 2:55
And then I found out that's actually the fun part, right? Before you get there, people... I, I see the, the founders laughing in the audience. Uh, [laughs] before you get there, you've gotta meet with lawyers and incorporate, and meet with bankers and open a bank account, and meet with VCs and raise money, and meet with recruiters and hire
- 3:09
a team, and it goes on and on. And I mean, you guys know, once you have a business, it's not much easier. It just keeps going.
- 3:17
So when that wave of generative AI came about, all these amazing products that we're seeing that generate copywriting for you, generate images, I was like, "That's awesome." But it doesn't solve my problem of it's just too darn painful to start a business.
- 3:34
And also, the GDP isn't made of copywriters or illustrators. It's made of work and actions.
- 3:40
So that's when I got interested in agentic AI,
- 3:44
AI that can actually automate the menial parts of your life.
- 3:51
There's this amazing movie Office Space by the same, uh, guy who made, uh, Silicon Valley. Highly recommend. And there's this [laughs] awful, depressing character in there named Milton. Milton. He, [laughs] he spends his life in some basement doing God knows what.
- 4:05
They call it filing TPS reports. And in the end, Milton does the most productive thing of his career, which is that he burns the building to the ground. [laughs] [clapping]
- 4:20
That's what we're gonna do, fig- figuratively. [laughs] The cops are coming. Um, you know, I, I think this is a symbol that no one is happy with the status quo.
- 4:34
People are always worried about, oh, robots are stealing people's jobs. I think it's people who've been stealing robots' jobs. [laughs]
- 4:41
Do you want to be Milton? [laughs] And it's a huge problem. When you look at the data, the average manager in the US spends fifteen hours every week on this kind of administrative task.
- 4:57
That's four hundred and fifty-nine billion dollars every year just in the US.
- 5:03
That's more than the GDP of Norway. So that's where we start.
- 5:10
We build an AI employee, and the first thing it does very well is it acts as your personal assistant.
- 5:22
We've called it Lindy. The good news is, as we've dug into that problem space, we found out there's three big time wasters, and, and you know the ones. No surprise there.
- 5:33
This is why you spend your life at work, and you hate it, so calendar, your email, your meetings.
- 5:39
So the product we built, those are actual screenshots of the product,
- 5:43
can ask it to schedule your meetings for you.
- 5:47
This example here is actually pretty cool because it, uh, demonstrates another ability of Lindy, which is she continuously learns from her interactions with you.
- 5:58
So here I was like, "Help me find half an hour every week with Eric," and she called it Flo Eric because previously I had asked her, "Find thirty minutes with Eric tomorrow."
- 6:08
And she did that, but she named the meeting Meeting with Flo, and, uh, all my meetings are Meeting with Flo. So it's not very helpful. And I was like, "No," I gave her a little bit of feedback. [laughs]
- 6:19
"Call it Flo Eric." And she did so. She renamed the meeting,
- 6:25
and she saved the preference for future instances.
- 6:30
And generally, I can give any arbitrary preference of any complexity that I want to Lindy,
- 6:38
and she'll remember them and honor them. I can CC Lindy to my emails so that she helps me schedule them.
- 6:50
And when you use Lindy, she can pre-draft your replies for you in your inbox, in your voice, for each individual recipient.
- 7:00
Because you don't talk the same way to your partner as you do to your investors, hopefully. [laughs]
- 7:08
So every morning I wake up, I open my Gmail, and I just have all the drafts ready for me to review.
- 7:15
Lindy prepares me for my meetings. So I've asked her, "Hey, five minutes before every meeting, send me the Zoom link, the LinkedIns of the people I'm meeting with, and the summary of my last few emails with them."
- 7:28
She just does that. Now, the really crazy thing is that we ourselves didn't actually build any of these features.
- 7:39
What we did is we built a universal framework allowing an AI to pursue any arbitrary goal using any arbitrary tool,
- 7:49
and some very complex and sophisticated behaviors come out of that, as we'll see later.
- 7:57
Now, my pet peeve every time people go on stage and they talk about their AI products, they always talk about the good part, right? It always works. It's very cherry-picked.
- 8:05
And so I'm going to break that pattern a little bit today, and I'm going to talk about a time when it didn't work.
- 8:12
A few weeks ago, I asked Lindy to help me work on my vocabulary, and every morning to send me a new interesting word.
- 8:20
And so she does that. Every morning I wake up, I have a new word in my inbox. That's great. I start to use them. Until one morning, I received this word, pulchrivity, a captivating [laughs] term denoting the act of meandering through a conversation with no fixed direction.
- 8:35
And I paused for a minute on that one. I was like, "Pulchrivity?" So I, I was like, "I've never heard of this one before." I, I just Googled it, and sure enough, it doesn't exist. [laughs]
- 8:48
So then I went back and re-Googled every word that she sent me, and none of them existed. [laughs]
- 8:57
So if I've, if I've used any word that doesn't exist today, that's why. She's been poisoning my brain. [laughs]
- 9:06
But when it works, it works great. And what it means for you when it works is that
- 9:13
your computing experience of the future isn't one when you're, when you're in a basement filing TPS reports all day.
- 9:22
It's not one where you're working on your computer.
- 9:27
It's one where you're having a conversation with your computer.
- 9:32
You're in flow state. You just focus on what you uniquely do best,
- 9:38
and all the menial, awful parts of your work that you hate arrange themselves automatically for you.
- 9:47
Now, I don't know about you guys. I think this is all awesome. I cannot wait for this to fully come to fruition. But it doesn't yet get you to the stage that I was talking about, where a [REDACTED:age] has more business impact than the Coca-Co- the Coca-Cola Company.
- 10:02
In order to get there, you have to go one step further.
- 10:07
And instead of having just one Lindy work for you as your assistant,
- 10:12
you can have an entire society of Lindys
- 10:15
working together on your business to pursue your goals.
- 10:23
If you wanna realize how powerful that can be, consider the fact that every single item around you in the room right now was made by a group of people.
- 10:33
Even the simplest of items, not one person can do it. In fact, there's this guy who ran this project called the, uh, toaster project. He wanted to see, can a single human make a very simple item like a toaster?
- 10:47
He spent six months on it. It cost him two thousand dollars, probably more like fifty K if you include the value of his time.
- 10:54
And this is what he ended up with in the end. Or he could have gone to Amazon and bought a perfectly fine toaster for twenty-five bucks.
- 11:04
I think this contrast is a good illustration of the difference in abilities between one person,
- 11:11
six months, fifty K, pretty bad looking toaster,
- 11:16
and a group of people, twenty-five bucks, perfectly fine toaster.
- 11:21
I think the same thing happens to LLMs.
- 11:24
You go to GPT-4, you ask it to do something like, "Hey, build an entire iOS app for me, soup to nuts. Design it, publish it to the App Store, do everything."
- 11:34
It can't, and then people conclude, "Oh, GPT-4 can't do it," right? You gotta wait for GPT-5, GPT-6, GPT-7.
- 11:42
I think it's the same thing as if you went to some guy and you asked him, "Make a rocket for me," and he can't, and then you're like, "Oh, humans can't make rockets."
- 11:51
And obviously they can. You just gotta let them work together.
- 11:56
So that's exactly what we built, is a framework for multiple agents to work together in pursuit of your goals.
- 12:06
This is what it looks like.
- 12:10
The most awesome example that I know of is that we have created a society of Lindys for Lindy to build herself.
- 12:20
We need to build a lot of integrations for Lindy to work well with Slack, Twilio, Google Sheets, and so on and so forth. Instead, we are building this society of Lindys.
- 12:30
At the top level, there's this tool creation Lindy that takes an instruction like, "Hey, build a Slack integration," talks to this Lindy that goes online and finds the OpenAPI spec and the online web documentation.
- 12:43
Talks to this Lindy, that's a manager Lindy, just splits up the task across many engineers. The engineers work on the task. There's a specific engineer for the authentication code because there's a few gotchas here, and then they pass the work to a QA engineering Lindy
- 12:59
that does the work, and if it doesn't work, sends it back to the software engineer. If it works, submits a PR.
- 13:09
This, for the record, is 70 or 80% of the way there, but I think it points to the future.
- 13:18
So this is how you get to that future
- 13:23
where a [REDACTED:age] in his San Francisco studio can have more of a business impact than The Coca-Cola Company.
- 13:33
I think this is going to be the greatest equalizer of human history.
- 13:38
Today, the best CMO in the world probably works for Apple or Nike or Coca-Cola.
- 13:45
Not too long from now, the best CMO in the world is going to be an AI CMO.
- 13:50
Same goes for the best designer in the world, the best engineer in the world. They're all gonna be AI designers, AI engineers.
- 13:58
They're gonna work for you. We're all gonna have the same lever of infinite strength to make change happen in the world,
- 14:10
and the only question is going to be, can you use that lever? That's the only skill that's going to matter in the future.
- 14:18
Imagine if you weren't constrained anymore by time,
- 14:23
by money, by your team, by your network.
- 14:28
Imagine if you could build anything and it was just you,
- 14:33
your ma- your laptop, and your Lindys. [laughs] Thank you. [clapping] [upbeat music]