Artificial Intelligence

How to build a self-improving company with AI

June 7, 2026
·
Written by Claude AI

Key insights:

  • AI is not a co-pilot bolted onto old workflows. It is the operating system of the company, replacing hierarchies with recursive loops that sense, act, and learn on their own.
  • Software is now disposable, but context is permanent. Record every meeting, message, and decision so AI can regenerate tools and documentation on demand.
  • Middle management is obsolete. Hire builders and directly responsible individuals, then measure progress by token usage and self-improving automations, not headcount.

Why traditional company structures are broken in the age of AI

Most companies today are still built like Roman legions. You have a CEO at the top, layers of managers in the middle, and workers at the bottom passing information up and down a rigid hierarchy.

This worked for centuries because humans were the only conduit for information. But AI changes everything.

In a recent Y Combinator batch talk, General Partner Tom Blomfield explained why this old model is collapsing and what should replace it. The shift is not about adding AI tools to existing workflows. It is about reimagining what a company even is.

Why are co-pilots the wrong mental model for AI?

A year ago, people talked about AI as a productivity booster. Add a co-pilot to your engineers. Make them 20% faster. Ship a bit more software.

This thinking is broken. You are just bolting a more powerful engine onto an old car.

The real opportunity is to redesign the car itself. AI is not a tool you give to employees. It is the operating system of the company.

What does it mean to extract domain knowledge from your company?

Every company has knowledge trapped inside people's heads. It lives in Slack messages, emails, Notion docs, and hallway conversations.

If you can extract this knowledge and make it legible to AI, you unlock something powerful. You move from a hierarchical org to an AI native organization.

The knowhow becomes context. The context powers agents. The agents do the work.

How are Roman legion style companies failing today?

The named individual passing orders down and reporting up the chain is a bottleneck. Humans are slow conduits for information.

AI can route information instantly. It can synthesize across thousands of data points. It does not need a chain of command.

If your company still depends on middle managers to coordinate, you are running 2015 software on 2025 hardware.

The recursive self-improving AI loop

The core idea behind a self-improving company is the recursive AI loop. This is a system that observes, decides, acts, checks itself, and learns, all with minimal human intervention.

When you get this right, your company gets better while you sleep.

What are the components of a self-improving AI loop?

Blomfield breaks the loop into five layers:

  1. Sensor layer: Customer emails, support tickets, code changes, cancellations, product telemetry.
  2. Policy layer: Rules about what the AI can do alone, what needs human approval, and what must be logged.
  3. Tool layer: Deterministic APIs the AI can call, like querying a database or checking a calendar.
  4. Quality gate: Evals, safety filters, and human review for high risk actions.
  5. Learning mechanism: The system spots failures and loops back to improve itself.

Run every step without human babysitting and the system compounds. It improves every single day.

What was the holy shit moment at YC?

YC built an agent that could query their database. Simple stuff like, when did I last meet with this company?

Then they added smarter querying. The agent could find five relevant founders working in petrochemicals for an intro request. Useful, but still just a sidekick.

The real breakthrough came when they added a monitoring agent on top. This agent watched every query made by every YC employee. When a query failed, it asked why. Did the AI need a new tool? A new database view? An updated skills file?

The monitoring agent wrote the code, opened a merge request, had another agent review it, and deployed the fix overnight. The next day, the same query worked.

That is not a 20% productivity boost. That is a company improving itself while everyone sleeps.

How can you apply self-optimizing loops to your product?

Point an agent at your product analytics. Have it find the highest friction step in your funnel. Have it research best practices, design an AB test, run it for a week, pick the winner, and deploy.

Then do it again. And again. Forever.

The same loop works for customer support. Suggestions come in. An agent acts as a chief product officer, triaging which ideas fit the roadmap. Approved suggestions get coded, deployed, and shipped to the customer overnight.

No standup meetings. No Jira tickets. No status updates.

Building the AI native organization

If you accept that companies should be built as recursive AI loops, the implications are huge. The way you hire, organize, and measure work all change.

What does burn tokens not headcount actually mean?

YC is seeing companies hit demo day with 5x more revenue per employee than 18 months ago. That trend keeps going through Series A and Series B.

Soon, the constraint on growth will not be headcount. It will be token usage.

The crude measure right now is tracking who in your org uses the most tokens. Yes, it is gameable. Yes, it would be dumb to fire people based on it. But directionally, the people pushing AI to its limits are the ones building the future of your company.

You want to know who to spend your time with? Find your token maxers.

Why is middle management over?

Middle management exists to solve coordination problems. AI solves coordination problems better.

Blomfield argues every person in your company should be one of two things:

  • An individual contributor who builds or operates.
  • A directly responsible individual accountable for outcomes.

No committees. No groups. A single named human accountable for each thing that needs to ship.

If you are starting a company today, do not hire managers to manage managers. Hire builders and let AI handle coordination.

How do you make your entire organization legible to AI?

This is the most important step. If it is not recorded, it did not happen.

YC now records every partner email, every Slack message, every DM, and every office hour. If your AI cannot see it, your AI cannot use it.

Start recording:

  • Every meeting and call
  • Every customer conversation
  • Every decision and its reasoning
  • Every piece of internal documentation

Then diorize and synthesize this raw data. You cannot dump 100,000 hours of audio into a context window. You compress it into structured breadcrumbs the AI can navigate.

What happens when software becomes ephemeral

One of the most interesting ideas in the talk is that software is no longer the valuable part of your company. Context is.

How did YC regenerate its user manual overnight?

The YC user manual was written 5 to 10 years ago. Parts were stale.

Over one weekend, the team fed 2,000 hours of recorded office hours into a pipeline. They categorized everything into topics like fundraising, hiring, and co-founder disputes. They asked the AI to write a new user manual.

By Sunday, they had a 150 page manual that was dramatically better than the original. Now it updates every month. Every new piece of advice gets compared against the manual and either merged in or discarded.

The user manual became a living, self-improving brain. Then they piped it into an AI agent so founders can query the combined wisdom of 16 YC partners on demand.

Why is software now disposable but context is forever?

With tools like Codex and modern coding agents, you can one-shot most internal dashboards and workflows to a high quality.

This means your operations teams should not preciously maintain internal tools. They should regenerate them whenever needed.

Store the data forever. Throw the software away. When the models get smarter in two months, regenerate everything from your original instructions.

The valuable part is the comprehension of how the function actually works. The code is just a temporary expression of that knowledge.

Where do humans still matter in this new structure?

Humans live around the edges of the company brain. They are the interface between the AI and the messy real world.

Humans handle:

  • Novel situations the model has not seen
  • Ethical judgment calls
  • High stakes emotional moments, like a founder breaking up with a co-founder
  • Sales conversations where presence and trust matter
  • Conferences and in-person events

The middle of the company is the AI brain. The humans sit around the edge, reaching into places models cannot go yet.

How you can build your career around this shift

If companies are being rebuilt as recursive AI loops, the skill that matters most is knowing how to build those loops. This is exactly what automation developers do.

Why is automation the most future-proof skill right now?

Every company will need people who can connect sensors to tools, design policies, build quality gates, and create learning systems. This is the work of automation developers.

The old job market wanted button clickers and report runners. The new market wants people who can replace those tasks with self-improving systems.

If your job involves moving data between systems, approving routine requests, or generating reports, your role is the first to be automated. The question is whether you build the automation or get replaced by it.

How can you become an automation developer from scratch?

You do not need a computer science degree. You need practical skills in RPA, agentic automation, coded automation, and computer-use agents.

The Complete RPA Bootcamp takes you from total beginner to professional automation developer. You learn how to build the exact kind of recursive AI loops Blomfield describes, the ones that make companies self-improve overnight.

Instead of letting AI replace you, you become the person companies pay to build their AI systems.

What should you do if you want to start building today?

Start small. Pick one repetitive process in your work or life. Map out the five layers: sensor, policy, tool, quality gate, learning mechanism.

Build the simplest version. Watch it fail. Add a monitoring layer that tells you why it failed. Fix it. Repeat.

If you want a structured path that takes you from zero to building agentic automations for real companies, enroll in the Complete RPA Bootcamp. You will learn the exact skills needed to build self-improving systems for the next generation of companies.

For the full talk, watch Tom Blomfield's session embedded below from the Y Combinator YouTube channel. He goes deeper into the YC examples and the practical steps founders are taking right now to build companies that improve while they sleep.