Author's Note

I wrote this manifesto because I believe the next generation of startups will be built around autonomous agents, not traditional apps.

Software is turning into something that does the job, and startups need to be built for that shift.

It's written for founders, and for the investors who back them.

Leo Celis

Edition Notes

Author: Leo Celis.

First published: June 8, 2025. Republished here as a living book.

This edition: v1.0.0, revised October 2026. Every quote was checked against a source, and quotes that could not be traced were cut or rewritten as the author's own words. Legal and market facts were updated to their state in October 2026.

Chapter 1: Every Revolution Needs Its Manifesto

In 2011, Marc Andreessen wrote that "software is eating the world."

Now software itself is on the menu. AI agents are eating software.

I've worked with startups since 2007, and I've used AI every single day since early 2022. What used to be a static software-as-a-service (SaaS) product is turning into something else: an AI-driven entity that acts on its own.

The people who fund startups see it too. In November 2024, the partners at Y Combinator argued on their Lightcone podcast that for every SaaS unicorn, you could picture a vertical AI unicorn alongside it.

So the next generation of founders won't build yet another app. They'll build autonomous digital teammates. Services that understand a goal, make the call, and act on the user's behalf.

And if you invest, the thing on the table changes with it. You're no longer looking at an app people log into. You're looking at an agent that does the work.

That sounds new. It isn't. People have been telling this story for more than 2,000 years.

The Bronze Guardian

The Greeks filled their world with gods and crafted heroes. They used them to explain nature, and to explain what people wanted.

A surprising number of those stories are about making powerful things. Even artificial life.

One of the best is Talos.

Talos was a man made of bronze. Hephaestus, the smith of the gods, made him to guard the island of Crete. Three times a day he ran around the island. Any ship that came too close got rocks thrown at it.

In the Argonautica, written in the 3rd century BCE, Jason and the Argonauts try to land on Crete. Talos stands in their way.

Medea brings him down. In one version she casts a spell, and he cuts his ankle on a rock. In another, she pulls out the bronze nail that seals his only vein. Either way, the ichor (the blood of the gods) runs out, and the guardian falls.

Here's the part that matters for founders. The Greeks imagined a machine with agency. Something that moved on its own and acted with a purpose. No prompt required.

The people who told those stories were the startup founders of the imagination.

The Myth-Maker Founder

Why go back to old myths in a book about startups? Because what we're building with AI agents taps into one of the oldest stories people tell.

The Myth-Maker Founder is the founder who ships two things at once: the agent, and the story that shows people how to live with it.

This isn't new either. The people who lead a change usually have to tell a new story to go with it.

Henry Ford didn't only sell the Model T. He sold the idea of everyday people driving, and a world built around cars.

Founders building AI agents are in the same spot, with a twist. Like the industrial builders, they're using a new technology to replace the old way of doing the work. Like the old storytellers, they have to give people a picture for it. Nobody has one yet for software that acts on its own.

You can already see the tech world reaching for the oldest pictures it has. In 2015, the engineer Anthony Levandowski registered a church built around the idea of an AI god. He shut it down in 2021, then revived it in 2023. But the instinct behind it is the one in the Talos story: a creation with its own agency and its own power.

These new entities are not tools. They are agents, software with agency and power, and they can land as helpers or trouble.

That depends on the founders who build them, and on the story they tell about them.

So writing a manifesto for agent-first startups isn't hype. It's part of the build.

Every revolution needs its manifesto.

Chapter 2: The Craftsman Shop and the Factory

So where are we in that history? There's a map for it. The one I use was drawn by an economist.

Carlota Perez studies technological revolutions. She counts five since 1771: the factory, steam and railways, steel and electricity, oil and the car, and then computers and telecom. Each one reshaped how people lived and worked.

Each one also starts the same way. In her book Technological Revolutions and Financial Capital, she writes that it begins with "a battle against the power of the old." The old companies, habits and institutions fight to stay on top. The new technology promises a better way.

Installation, Then Deployment

Perez splits every revolution into two periods.

First comes the installation period. The new technology is, in her words, "a small fact and a big promise." It has to prove itself. Money floods in, a lot of it on hope. People experiment, and plenty of it fails. (Sound familiar? It should. We saw it in the dot-com boom, and we're seeing it again with AI.)

Then comes a turning point. Often it's a crash. Perez reads the dot-com bust as exactly that turning point for the computer age.

After it comes the deployment period. The new way has beaten the old one. It becomes normal, and the gains spread to everyone.

AI is deep in its installation period. A small fact and a big promise, already strong enough to push the old way out.

And the old way, in software, has a name. For the last two decades it has been software as a service (SaaS). Startups built on agents are the new challengers, with a different way of creating value.

The friction is the same as every time before. Older companies resist. Skeptics wait. Founders push forward.

So breaking with the SaaS playbook isn't a jump into the unknown. It's a pattern people have lived through five times.

The Factory Next Door

The closest match to what's happening now is the Industrial Revolution of the 18th and 19th centuries.

Before it, most things were made by hand, at home or in small workshops. After it, they were made by machines in factories. Water wheels, then steam engines, then electricity changed how goods were made, and how people lived. Whole villages moved to factory towns.

The people who drove that change walked away from the old way of making things.

Take Richard Arkwright. Instead of one craftsman making one item at a time, he built water-powered mills that spun cotton thread nonstop. He built his mill at Cromford in 1771. Perez treats it as the start of the first revolution, the big bang of the factory age.

It was a small fact and a big promise, like today's AI prototypes.

Arkwright and the mill owners after him were the startup founders of their time. They proved a new model while plenty of people distrusted it.

We forget how deep that resistance went. Skilled textile workers feared for their work. The Luddites went further and smashed the new machines.

But the math of the factory won. Whoever ran power looms and steam engines could make far more, far cheaper, than whoever stayed with the hand loom. The battle against the power of the old was fought, and the factories won it, across Britain and then the world.

Then each new wave rode on the one before: steam and railways, then steel and electricity, then oil and the car. Mass-produced goods. Higher living standards. A growing middle class.

The leaders across those waves were the founders who went all in on machines: Carnegie in steel, Rockefeller in oil, Edison in electricity, Ford in cars. The ones who held on to the old way ended up running shops that looked quaint, then obsolete.

The Craftsman-Shop Trap

Now look at software through the same lens.

If the classic SaaS startup is the craftsman shop of the digital age, then a startup built on AI agents is the new factory.

The Craftsman-Shop Trap is building a better tool for people to operate while the factory next door does the work itself.

Machines took over and scaled up the manual work of the 19th century. AI agents can take over and scale up the thinking work and the services of the 21st.

Picture a founder still building a classic SaaS product. A user logs in and does the work by hand with the tools. That founder risks ending up like a hand-loom weaver in the 1820s, watching the power loom outrun him.

The factory didn't end the need for people. It changed what their work was multiplied by.

In the industrial age, human muscle was multiplied by machine power. In this one, what people bring is multiplied by agents that do the thinking work: the trust others place in them, and the knowledge their startup builds.

And if you invest, Perez's map is your map too. The frenzy belongs to the installation period. The shakeout is the turning point after it. What matters is who is building the factory, not who is polishing the shop.

The lesson from the Industrial Revolution is simple: to lead the next era, take up the new way of making things.

For startups, that means AI agents that do the work, not only SaaS tools that people operate. It means building the factory, not the shop.

The change is simple: agents do the work, and the buyer gets the outcome.

Chapter 3: From SaaS to Digital Entities

I want to start with a hammer.

For two decades, software as a service (SaaS) sold tools. A vendor built the software, put it in the cloud, and rented it out by the month. The customer logged in and did the work.

An accounting tool is the clean example. A startup pays for the subscription. Then someone on the team still has to do the accounting, inside that tool, by hand.

The vendor sold the hammer. The customer swung it.

The Autonomous Carpenter

Now picture a different deal. The same startup doesn't rent accounting software. It signs up for an AI accountant. The agent closes the books, chases the receipts, and flags what looks wrong. A person checks the result.

The Autonomous Carpenter is software that does the job the tool was bought for, and sells the finished work instead of the tool.

In the old world, the vendor provided the hammer. In the new one, the vendor provides the carpenter.

It's the difference between selling someone a car and giving them a driver.

For the customer, the value moves up a level. They pay for an outcome, not a process. For the startup that builds the agent, the responsibility goes up too. If the books are wrong, it's their problem now, not the user's. The control goes up with it. The startup decides how the work gets done, not each user.

Why the Prize Is So Large

Here's why investors are paying attention.

Software was always a slice of the economy. Services were the much bigger slice: legal work, accounting, support, sales, finance. Those are paid as salaries and outsourcing, not licenses.

The venture firm Foundation Capital puts the opportunity for agents that do that work at $4.6 trillion. That's their estimate, and estimates move. The direction doesn't. Agents let software startups go after the services budget, not just the software budget.

And that's the floor. The services budget only counts work someone already pays people to do. The bigger prize is the work nobody could afford to staff.

So these startups will look different. They may have far fewer people for the revenue they bring in. Their product may feel more like a hire than an app. You'd bring them on the way you'd bring on a firm.

And they'll be judged on results. Did the agent close more deals? Did it resolve the ticket? Did the books balance? Logins and seats stop being the score.

If you invest, that changes the question you ask in the pitch. Not "how many users?" but "how much work did the agent finish, and how often was it right?"

From Workflows to Goals

Letting go of the SaaS mindset is hard. I grew up in it too.

SaaS taught founders to win on distribution, engagement and features. All good things. But all inside one frame: the user does the work with our tool.

The agent frame asks a different question. What whole problem could we solve if our software were smart enough to handle it?

Take project management. A SaaS tool gives people features to manage their tasks. An AI project manager manages the project. It assigns the work, follows up, changes priorities, and pulls in a person only when it needs one.

That's the move from user workflows to user goals.

Interface Versus Agent

A classic SaaS app is a set of features behind a screen. Every feature waits for a person to click.

An agent works more like a digital employee. It reads the context, learns from the data, and adjusts as it goes. It doesn't wait for the next click.

The people running the biggest software platforms are saying the same thing. On the BG2 podcast, in December 2024, Microsoft's Satya Nadella described most enterprise apps as databases with rules on top, and said those rules are moving to AI agents. At Dreamforce, in September 2024, Salesforce's Marc Benioff called agents the third wave of AI.

And in January 2025, OpenAI's Sam Altman wrote: "in 2025, we may see the first AI agents 'join the workforce' and materially change the output of companies."

The First Carpenters

You can already see the early versions.

GitHub Copilot started by suggesting the next line of code. Now it writes whole changes for developers. Support agents close whole tickets, not just suggest replies. AutoGPT showed, in rough form, an agent that breaks a goal into steps and works through them on its own.

Then came startups that put the agent at the center from day one.

In March 2024, Cognition revealed Devin and billed it as "the first AI software engineer." Devin plans, writes and debugs code. It's pitched as a teammate, not a tool.

Lindy bills itself as a way to "Hire an AI Employee That Gets Work Done." Its agents draft email, schedule meetings and answer customer tickets.

These are early and narrow. Some of the claims around them are bigger than the results. That's normal for an installation period.

Software shouldn't sit there waiting for input. It should take the job and do the work.

A startup that sells finished work can outrun one that sells a tool and leaves the work to the customer. The economics point one way: speed, scale and cost all favor the carpenter.

The startup that owns the work beats the one that rents a tool.

Chapter 4: From Products to Agents

Once the agent does the work, the product changes. It stops helping with the job and starts doing the job.

I find it easiest to see in pairs: the old SaaS way next to the new agent way. Five pairs change at the same time. That's the point. You don't get to change one and keep the other four.

The Five Shifts are the five things that flip together when a product becomes an agent: tool to teammate, features to outcomes, static to evolving, hiring to deploying, and efficiency to autonomy.

Shift 1: From Tool to Teammate

Old: people use software through screens. They click buttons, fill forms and read dashboards. The software helps, but the person runs every step.

New: people talk to an agent the way they'd hand work to a colleague. You say what outcome you need. The agent works out the steps.

So the agent stops being a tool and starts being a collaborator. It takes the busywork and even makes suggestions.

In practice, every user becomes a manager. They assign tasks to an agent instead of doing the tasks themselves.

Shift 2: From Features to Outcomes

Old: startups shipped features. Each feature solved part of the problem. The user strung them together to get a result. The product was a toolkit, and it only paid off if the user worked it well.

New: startups ship agents that deliver the outcome. Not a long list of features. One agent that solves the whole problem.

Think about travel. A classic travel app gives you forms to search flights and hotels. An agent plans the whole trip from one request.

Inside the startup, the work changes too. You stop writing code for every step. You set the goal, give the agent the tools and context, and let it find the path.

Shift 3: From Static to Evolving

Old: SaaS got better through releases. Between releases, the software did exactly what it was coded to do. A new use case meant people had to redesign the product.

New: a well-built agent learns from each job, through memory and feedback, and adjusts. It improves with new data and new experience. It doesn't go stale on a shelf between versions.

The release cycle gives way to learning while it runs. The team's job becomes guiding that learning: giving feedback, setting goals and limits, and upgrading the model when a better one comes out.

Shift 4: From Hiring Humans to Deploying Agents

Old: growth meant hiring. More customers meant more sales reps, more support staff, more operations people. Headcount grew with ambition.

New: an agent-first startup grows by running more agents. It can run hundreds of copies at once, around the clock, at a small cost per copy. When demand spikes, you don't hire and train a dozen people. You add capacity.

Investors have noticed what that means for the market. José del Barrio and Carmen Sánchez of the venture firm Samaipata wrote in November 2024 that "AI agents will significantly expand software's total addressable markets by tapping into the largest budget of all: payroll."

So scale stops tracking headcount. My expectation is that agent-first startups will run many more agents than people. The people will act as managers and coaches. They train the agents, check their work and steer them, instead of doing the front-line work by hand.

In the old model, software was a tool for the team. In the new one, the software is a big part of the team.

Shift 5: From Efficiency to Autonomy

Old: SaaS sold efficiency. "We save you time on X." "We help you do Y better." But the user still did the thing. The value stopped at the limit of the user's time and attention.

New: an agent takes the whole task. It doesn't make you faster. It takes the work off your plate.

It answers the support request at 3 a.m. It works through a million documents. It doesn't get tired.

That moves software out of the tools budget and toward the labor budget. Take sales. The old pitch was a CRM and a sales team to work it. The new pitch is an AI sales associate that finds leads, sends the emails and books the meetings.

Outcome, not effort, becomes the unit of value.

Why They Move Together

Each shift needs the others. An agent that delivers outcomes but waits for clicks is still a tool. An agent that learns but is priced like a seat is still SaaS.

That's why a bolt-on rarely works. Adding a chat box to a classic product changes one shift and leaves four.

If you invest, the five shifts work as a checklist. Ask where the startup sits on each one. Most pitches will be new on one or two and old on the rest.

If agents do the work, the startup itself has to be built differently. The whole company has to be shaped around software that does the work.

Chapter 5: Principles for an Agent-First Startup

Principles come before tactics. Tactics change every quarter. Principles decide which tactics you even try.

I keep coming back to six. None of them is about a model or a framework. They're about how you think while you build.

The Hire Test is treating your agent the way you'd treat a new hire: give it a job, judge it on results, talk to it, coach it, build the team around it, and charge for the work it does.

Principle 1: Build Agents, Not Apps

Picture your product as a team member, not an application.

Ask one question. If my product were a person I hired for this problem, what would their job be? Then build the agent to do that job.

Take legal tech. Don't build a document manager. Build an AI paralegal that reads, analyzes and drafts on its own.

That moves you off the feature list. The agent has a purpose, and the purpose is an outcome.

An agent also lives in its domain. It keeps running, takes in new work and remembers what came before. Give it a role, like financial advisor or project manager, and you're delivering a service, not shipping software.

Principle 2: Chase Outcomes

Make the user's result your north star.

Don't ask what feature to build next. Ask what bigger goal the agent can take on next.

Measure what the agent gets done: tasks finished, problems solved, value created. Clicks on a feature tell you less.

Start design with a conversation. What would a user ask the agent to do? How would you know it succeeded? Then work backward to what the agent needs to deliver that.

When the agent falls short, make the agent better. Don't add buttons for the user to press.

The user says what. The agent figures out how.

Your job as founder is to keep getting better at the how.

Principle 3: Talk Instead of Click

People instruct each other in plain language. So the natural way to instruct an agent is a conversation. That can be a chat box, a voice command or a short form.

It goes past the chat window. The old interface showed people how to do something. The new one lets them say what they want done.

Think about how the agent reports back. How it asks when it's unsure. How it shows the result.

Devin, from Cognition, is a good example. In March 2024 Cognition wrote that "Devin reports on its progress in real time, accepts feedback, and works together with you through design choices as needed."

That loop builds trust. The user stays informed without managing every step.

Design dialogues, not dashboards. Working with your product should feel like working with a capable colleague who speaks the user's language.

Principle 4: Coach the Agent

In an agent startup, building the product is a partnership between people and AI.

You'll manage your agents the way you manage staff. You give them feedback, new knowledge and new duties over time.

Test the agent on real work, often. See where it fails. Fix it, and test again.

Your team ships code, and it also shapes behavior. It reads what the agent decided, finds where it went wrong and corrects course.

The agent's performance is the product. Watch how it behaves out in the world.

Plan for memory too. The agent should keep what it learned from past work, inside privacy limits, so it doesn't repeat mistakes and it serves each user better.

That turns quality checks into ongoing training. It's how the product stays ahead of what users expect and what rivals ship.

Principle 5: Build the Team Around the Agent

People and agents work together. Sometimes the agent leads.

It starts inside. Have your team hand their own routine work to agents. Then they know what it's like to be the customer.

Hire people who want to direct agents, not just write code. Domain knowledge matters more than it used to. It's what your agents learn from.

Keep the ratio of agents to people high as you grow. Not to cut people for its own sake. To stay small and keep people on the work agents can't do yet.

Every person you hire should multiply their impact through the agents they run.

So make it a rule. When a new problem or workflow shows up, ask "can an agent handle this?" before you hire for it.

Celebrate when an agent hits a milestone, the same way you would for a person. It's part of the team.

And set the ethics on day one. Autonomy brings responsibility. Train the team to watch for bias, errors and side effects, and to own them. As agents act more on their own, they have to stay aligned with what users want and what the startup stands for.

Principle 6: Sell the Result

Agent startups get to redraw how they charge.

If the agent delivers the outcome, price on the outcome. Charge by tasks completed, or take a fee when it succeeds.

Tie your revenue to what the agent actually delivers. Customers trust that.

Pick problems where an agent can be ten times better than how they're solved today. The best targets are markets software never reached, because the work was done by people.

Pete Flint and Anna Piñol of the venture firm NFX put it this way in April 2024: "Labor and software are fusing into one massive market."

So frame the vision as a new way to solve the problem, not better software for it.

The old pitch: a platform to manage your social media marketing. The new pitch: an AI marketer that runs your social media accounts for you. One sells a tool. The other takes the hassle away.

Build your message on that promise. Then price on the value the agent creates, not on the savings. Savings top out at what the old work used to cost. That's the $99 Savings Trap.

If you invest, two questions sort the field. Does the pricing move with the results? And did the founders ask "can an agent do this?" before they hired?

An agent that acts for people has to answer to the law and to the people it serves.

Chapter 6: Declare the Machine

The bronze guardian Talos had one vein, sealed by a single bronze nail. Pull the nail, and the guardian fell.

I think about that nail a lot. Today's agents are far more capable than Talos, and they have limits too. Laws, oversight rules and duties of accountability decide what a responsible agent has to be.

Ignore them and you build a marvel that no regulator will allow and no serious buyer will touch.

So the job isn't only to build what the agent can do. It's to build in what it must and must not do. That stopped being optional.

Declare the Machine is the duty to tell people, clearly and every time, that they are dealing with an AI, and to build the agent so a human can review it, stop it and answer for it.

Privacy and Human Review

Start with Europe's General Data Protection Regulation, the GDPR. It reaches startups anywhere in the world that offer services to people in the EU or track what they do.

Article 22 gives people the right not to be subject to a decision made solely by automated processing when it has legal or similarly significant effects on them. Where such decisions are allowed, people keep safeguards. One is "the right to obtain human intervention." They can also give their view and contest the decision.

For an agent startup, that means a clear path to a human. Hiring, lending and other decisions that change people's lives need one.

The model itself can fall under the law too. In December 2024 the European Data Protection Board said that AI models trained with personal data "cannot, in all cases, be considered anonymous." Each case is judged on its own.

The Board later turned that warning into operating guidance. In July 2026 the EDPB adopted guidelines on anonymisation and web scraping for generative AI that build on its AI-models opinion, clarifying when data is truly anonymous and how legitimate interest can apply to AI training.

That brings in Article 25: data protection by design and by default. You plan the safeguards before you ship, not after.

The highest GDPR fines reach €20 million or 4% of worldwide annual turnover, whichever is higher.

Rules by Risk: The EU AI Act

Europe's AI Act, Regulation 2024/1689, sorts AI by risk. The higher the risk, the heavier the duties.

The high-risk list, Annex III, names eight areas where agents are already being built: biometrics; critical infrastructure; education and vocational training; hiring, managing workers and access to self-employment; access to essential private and public services and benefits, including credit scoring; law enforcement; migration, asylum and border control; and justice and democratic processes.

The dates moved. A 2026 amendment pushed the high-risk duties for Annex III systems to December 2, 2027. AI built into regulated products follows on August 2, 2028.

What those systems must have is already written down:

1. Human oversight. People must be able to oversee the system while it runs, and to stop it with "a 'stop' button or a similar procedure."

2. A named human with real authority. The organization using the system has to give oversight to people with the competence, training and authority to act.

3. Logs. The automatically generated logs must be kept for at least six months.

Fines come in tiers. Banned AI practices can cost up to €35 million or 7% of worldwide annual turnover. Breaking other duties, including the AI disclosure rule, can cost up to €15 million or 3%. For startups and other small firms, the cap is whichever of the two is lower.

If your agent works in one of those eight areas, whether it acts or only advises, assume you're in scope. Building it in early is cheaper than retrofitting it.

Tell People It's an AI

One rule that reaches almost every agent already applies. Since August 2, 2026, Article 50 of the AI Act requires that people who interact directly with an AI system "are informed that they are interacting with an AI system," unless it's obvious.

California got there first. Its bot disclosure law, SB 1001, has applied since July 2019. It's unlawful to use a bot online to mislead a person about its artificial identity in order to drive a sale or sway a vote. A bot that discloses it's a bot is in the clear. The disclosure has to be "clear, conspicuous" and built to actually inform.

No buried footnotes. No ambiguity.

Truth in the interface is the first rule of an agent that speaks for anyone.

Voluntary Codes

Not every standard is law. Three are worth knowing anyway.

The OECD AI Principles were adopted in 2019 and updated in May 2024. The NIST AI Risk Management Framework, released in January 2023, is voluntary by design. ISO/IEC 42001:2023 sets out a management system standard for AI.

My view: founders who adopt these early sell faster into regulated buyers and rewrite less later. Trust is a feature in markets where the stakes are high.

Ethics by Default

Ethics can't be bolted on at the end. Build it in from the first line:

1. Declare the machine. Whether the user is in Fresno or Brussels, the interface says: AI at work.

2. Keep a hand on the stop. Every agent ships with a stop that a named person has the authority and the duty to use.

3. Record the lineage. Keep a record of how the model was built, trained and deployed, and what it decided.

4. Support human override. Design the agent as a deputy, not an autocrat. It can be wrong, reviewed and interrupted.

5. Audit all the time. What the agent does in the wild is the real product. Watch the edge cases and correct course early.

If you invest, these five are diligence questions. Ask to see the disclosure, the stop, the logs and who owns the override. A team that can't show them hasn't built them.

Chapter 7: Join the Agentic Revolution

The old SaaS playbook worked. It built a generation of great startups. I think it's reaching its limit.

When agents can learn and do this much, holding on to the old model is like insisting on the horse while the first cars roll past. Not wrong yesterday. Wrong now.

For founders and builders, the agent-first model is no longer optional theory. It is the model.

The Agentic Startup is a startup whose core team can be a founder, maybe a co-founder, and agents doing the work, where the product is the work done, not software someone else has to operate.

It Won't Be Easy

Paradigm shifts never are.

It takes nerve to drop old product assumptions. To build on technology that's still moving. To teach a market a new way of buying.

But that's where startups win. At the edge of what's next, pulling the future forward.

The reward isn't only revenue. It's the chance to change how the game is played.

What It Looks Like

Agent startups aren't iterating on old ideas. They're building toward a world where every startup can field a crew of tireless digital workers. Where people have personal agents that handle what they need. Where growth is limited by imagination, not headcount.

Founders launch with agents as the first hires, running the work from day one. Small teams set the goals and let agents test and tune the answer in hours. Customers say "I have an AI for that" about everything from shopping to booking a plumber.

That's my picture, not a forecast. But some of it is already in the data. Carta, which tracks startups on its equity platform, reports that the share of new startups with a single founder rose from about a quarter in 2019 to more than a third by mid-2025.

That trend kept going. In March 2026, Carta reported that about 36% of startups founded on Carta in 2025 had a single founder.

Fewer founders per startup. More work done by software. The two lines point the same way.

The Startup as an Entity

Let's stop thinking of startups as apps or websites. Think of them as entities. Goal-driven. Getting better with every job. Ready to serve each user in their own way.

The last eras were mobile-first and cloud-first. This one is agent-first.

It means daring to build a startup where the core team might be you, a co-founder and a group of agents doing the work a large team used to do.

It means trusting that with careful training, alignment and oversight, those agents can match people on speed, consistency and scale, and then pass them. The people keep what agents can't hold yet: the trust others place in them, and the knowledge only their startup has.

And it means keeping the guardrails from the last chapter. An agent that acts for people has to be declared, overseen and answerable.

The question isn't whether this happens. It's already happening. The question is whether you watch it or lead it.

Back to Crete

This book started with a bronze guardian. Talos ran his circuit around Crete three times a day, and nobody had to tell him to.

The Greeks only imagined him. You get to build him.

Tell the story that teaches people how to live with your agent. That is the Myth Maker Founder job. Build the factory, not the craftsman shop. Sell the carpenter, not the hammer. Make the agent's role obvious, and its boundaries visible. Declare the machine.

Talos had one weak point, and Medea found it. Your agent should have a stop, and a person whose job is to hold it.

The Invitation

Build something bold. An agent that pushes past what we thought software could do.

Do that, and you won't only build a startup that defines its category. You'll move the whole market, and give others the nerve to follow. That's how a manifesto becomes a movement.

The startups that win will be small teams whose agents do the work, priced on results, with the guardrails already built. Their ratio of work done to people employed will look strange. Their edge is what those agents learn on every job that a rival can't copy.