Video Title: How to Build a Self-Improving Company with AI

Video Author: YC Root Access

Translation: Peggy

Editor's Note: In this latest YC batch talk, YC General Partner Tom Blomfield discusses not "how to use AI to improve employee efficiency," but a more fundamental question: when AI is no longer just a Copilot, but can perceive, make decisions, call tools, receive feedback, and self-correct, how should the company itself be redesigned?

Tom's core thesis is that traditional companies still operate like "Roman legions": information flows upward through hierarchies, and commands flow downward through management chains. But AI is breaking this organizational assumption. What truly matters is not getting engineers to write 20% more code, but extracting the business knowledge scattered across emails, Slack, meetings, documents, and people's minds, turning it into an organizational context that AI can read, call upon, and iterate on.

In his view, the future AI-native company will be composed of a series of recursive, self-improving AI loops: the system perceives external changes from customer emails, support tickets, and product data, executes decisions through rule layers, tool layers, and quality gates, and then automatically learns and corrects based on results. YC is already experimenting with a similar mechanism: the agent doesn't just answer questions; it monitors which queries fail, determines whether a new tool, database, or index is needed, and automatically commits, reviews, merges, and deploys code. In other words, the company can continue to optimize itself while the founder sleeps.

This also means AI's impact on companies won't stop at the tool level; it will further change organizational structures. Tom proposes "burn tokens, not headcount" – the bottleneck for future startups may no longer be the number of employees, but token usage, the quality of business context, and the readability of organizational knowledge. The coordination functions performed by middle management will be largely replaced by AI, while ICs, directly responsible individuals, and human roles capable of handling high-risk judgments in the real world will become more important.

The most noteworthy point is not that AI makes companies more efficient, but that it is changing the very organizational form of "the company" itself. When software can be generated on the fly, processes can be automatically improved, and experience can be continuously沉淀 (沉淀) into a company brain, what founders truly need to build may no longer be a clear hierarchical team, but an intelligent system capable of continuous learning and self-optimization.

Below is the original text:

Rewriting How Things Work: Companies Should No Longer Operate Like Roman Legions

This part is somewhat based on a previous talk by Diana. The video from the weekend is already online and is fantastic. Also, Jack Dorsey posted some tweets about two or three weeks ago that I found very interesting, so I "borrowed" quite a few of his ideas and incorporated them into this session.

This talk will be somewhat conceptual and high-level, mainly discussing how we should rethink building companies.

The Roman legion was designed, in essence, to project power from the center of Rome across two continents, all the way to Hadrian's Wall near Scotland. It relied on a nested hierarchy, with a stable span of control at each level. Each level had a clear person in charge, responsible for passing commands down and relaying information back up.

If you look at most companies today, you'll find they still operate like a Roman legion: people are the channels through which information flows up and down. What struck me in Jack Dorsey's thread of tweets is that we have always assumed hierarchical organization is the best way to organize economic units of value. But I believe AI is fundamentally breaking this assumption.

A year ago, if you asked people what AI is for, they would usually talk about "productivity": for example, Copilot makes engineers 20% more efficient, integrating Copilot into workflows helps teams deliver more software. But I think this is a problematic way of understanding it. It's like putting a more powerful engine on an old way of working. What's truly worth thinking about is not how to add an AI tool to an old organization, but reimagining what the company itself is and how it should operate.

For example, take what Garry just talked about. I truly believe he alone can now produce as much code as an entire engineering team. What really keeps me thinking is how to extract the domain knowledge inside a company and define it as context, skill sets, or whatever you want to call it.

This domain knowledge, business knowledge, know-how – it was originally scattered in people's minds, Slack messages, emails, and Notion documents. This information collectively defines how your company operates. Once you can make this knowledge clear and readable, you can move from a hierarchical organization to an intelligent organization driven by AI-native software.

Making the Company Better While You Sleep: How AI Loops Automatically Discover, Fix, and Deploy

AI is not something attached to the side of a company. It's not just a tool for engineers to improve efficiency. I believe we can reimagine a company as a set of recursive, self-improving AI loops. This point is very important because once a company reaches this stage, it will continuously self-optimize even while you sleep.

For example.

Diana also mentioned this AI loop in her talk. It starts with a "sensor layer." This term sounds fancy, but it can be quite simple: emails from customers, support tickets, code changes, user cancellations, product telemetry data – these are all sensor data used to gather information from the outside world.

Then comes the policy or decision layer, i.e., the rules: what the AI can do, what things must request human permission, what operations must be logged. Below that is the tool layer, which is somewhat like what Garry called skills and code – essentially deterministic APIs, such as querying databases, checking calendars, etc., a set of tools the AI can call.

Next is the quality gate, like the deterministic checks, security filters Eva mentioned, and human review for high-risk matters. Finally, there's the learning mechanism: the system interacts with the real world, discovers what isn't working, and feeds that feedback back to the start of the loop.

If each step can run without, or with minimal, human intervention, the system will get better and better while you sleep.

I can give you some examples of what we are actually running now. Initially, we built an agent that you could ask questions, and it had some deterministic tools to query our database. For example, a simple question: When was the last time I had office hours with this company?

Later, it got smarter. For instance, if I'm doing office hours with a company and they need to know someone in the petrochemical industry, the system can query the database in different ways, combine it with RAG and other methods, find five relevant founders, and recommend them to you.

But this is still just a sidekick, a helper agent. It's still the way AI was used last year: AI makes me, as a group partner, more efficient, boosting my work efficiency by 20% or 30%.

What really gave me the "aha moment" was when we added a monitoring agent on top of this system. It looks at every single query made by every YC employee, judges which queries succeeded and which failed. Then it asks: Why did it fail? What would it take to make this query succeed? Do we need a new deterministic tool? Do we need to update a skills file? Do we need a new database? A new index?

These things now happen automatically at night. It writes code, submits a merge request to YC's codebase, has another agent review it, then merges and deploys it. So the next day, when a human comes and asks the same question, the query can succeed.

For me, that was the crucial moment. It's not just about making a human 20% or 30% more valuable. It's the AI completing the loop itself and finding a way to self-improve.

I believe that if you can identify which parts of your company can operate this way, and minimize the human execution and supervision role in them, then you can pour tokens into that problem, and the company itself will continuously get better.

There are many other examples. For instance, if you have product analytics data, you can have an agent analyze the product data to find the biggest friction points in the sales funnel. It can research best practices, set up an A/B test, run it for a week, pick the best performing variant, and deploy it live.

This happens over and over again. Your product will have a self-optimizing product loop.

The same goes for customer support. Customer suggestions keep coming in. You can use an agent for triage. This agent acts somewhat like your Chief Product Officer and Chief Technology Officer, making judgments: we don't want to do this suggestion, discard it; but this other suggestion aligns with our roadmap and can be done tonight. Then it writes code, deploys it, goes live, and delivers it directly to the customer, all without human intervention.

So, if you can view every part of the company as a self-improving recursive AI loop, it becomes something very different from a "Roman legionary" hierarchical company.

Burn Tokens, Not Headcount: AI-Native Companies Will Reshape Organizational Structure

So, what does it mean if you want to do this?

The first point is: consume tokens, not pile up headcount. We are already seeing that by Demo Day, the revenue per employee for many companies is about 5 times higher than it was 18 months ago. I think this trend will continue into Series A and B rounds. Soon, your real constraint won't be the number of employees, but token usage.

The crudest way to measure this now is to look at each person's token usage. Of course, this metric sounds stupid in extreme cases and is easily gameable. But directionally, I think it's right. We are in a phase of exploring "what is actually possible," so everyone should experiment maximally to see what this crazy new intelligence can do.

Once you make it a leaderboard and link promotions or terminations to this metric, it will certainly be gamed and become distorted. But directionally, figuring out who in the organization is leveraging tokens to the fullest and who isn't is indeed a way to decide which employees you should spend your time on.

I believe middle management is over. At least for this kind of coordination problem, I don't think middle management is needed anymore; AI should handle this.

For me, there are two important roles in the future. Jack Dorsey mentioned three, but I didn't like the third, so I removed it. I think the two really important roles are: everyone must become an IC, i.e., individual contributor, builder, operator. And crucially, there must be a directly responsible individual. For anything to move forward, there needs to be a named person clearly responsible, not a committee, not a group.

I believe companies can be built entirely on ICs. Middle management is truly over. And building a self-improving company is this vision.

By the way, I think everyone is still at the very frontier of this. I'm also very curious to see how far you've gotten. It feels like everyone is still exploring the boundaries. I'm not sure if there's someone who has built a truly self-improving company across every function yet. Maybe I'm wrong, and you can prove me wrong.

If it were me, what would I do first?

The first very important thing is to make the entire organization readable and understandable to AI. What does that mean? It means you must record everything.

Simply put, for all our partners' emails now, if you send an email to a YC partner, that email goes into the YC database. Every Slack message, every DM, every office hours session – we've started recording all of them for the past three or four months. Everything that happens, as long as it's recorded, has happened for the AI; if it's not recorded, it hasn't happened for your intelligent system.

Earlier I was chatting with some founders here, and we talked about a lot of good content regarding their companies. Every time we chatted, I thought, I really should have recorded this conversation. Because someone just needed an introduction from me, and now I can't even remember who the introduction was for. I said yes at the time, then told him to email me later because I knew I would forget, as I had 20 more people to talk to next.

So, this might require phones, recording devices, smart glasses, or putting microphones in every room. In short, everything needs to be recorded so the AI can read it.

Then, as Garry said, you also need speaker diarization and summarization. You can't just throw 100,000 hours of recording into a context window. You must process them, aggregate, compress, distill them into important parts, and leave some clues for the AI.

For example: How many of you have read the YC user manual? Hopefully, everyone in this room has opened it at least once. It's okay. Most of that manual was written five to ten years ago, it's a bit outdated.

Last weekend, Harsh had a sudden thought: Since we've accumulated about 2000 hours of office hours recordings over the past three months, why not regenerate the user manual?

So, you can give the system a set of instructions, first process, compress, and synthesize the recordings, then categorize them by topic like fundraising, hiring, co-founder disputes, etc., and then have it write a new version of the user manual. By the end of the weekend, he had generated a 150-page user manual that was significantly better than the existing version.

More importantly, now we can update it monthly. So our user manual has become a self-improving system. Every new piece of advice is compared to the existing manual, either absorbed or discarded. Thus, the user manual becomes a continuously updated, living brain, containing the advice we give founders every week.

Of course, it doesn't stop at the user manual level. You can input this as context for the AI agents. Suddenly, you can ask a super-intelligent AI a question and get the combined wisdom of 16 YC partners. But the precondition is that this knowledge must be readable by the AI. So you must record everything.

The second point is similar: If something can create a self-improving artifact that can be read by AI, keep it; if not, discard it.

The third point is that every function should be able to generate its own software. In the past, we might have said "dashboard," but now it's not just dashboards; it's software generated on demand. Codex 5.5 is now good enough that for most simple internal software and dashboards, you can generate them to a fairly high quality in one go. I tried it with some of our internal stuff over the weekend, and the results were truly incredible.

So, all internal operations teams should sit on top of this layer: possessing an intelligent understanding of the business, and then generating their own dashboards and workflows.

And I would treat this software as completely disposable. What should be very carefully preserved is the data. As Garry said, he saves all his emails as Markdown, never discarding anything. But the software itself is ephemeral, temporary. You can generate it, and you can regenerate it.

What is truly valuable is the understanding of the business in people's minds: how this function operates, how we run a YC event, etc. As for the software actually used to run the event, you can generate one for this event, use it, and then discard it. A month or two later, when the model gets smarter, you throw away the old software, give it the original instructions again, and generate a new version of the software.

So, I believe what is valuable is the business context and skills. The software built on top of them is temporary.

So, in this world, what is the role of humans?

I think what we are discussing is essentially a "company brain." I know many people in this room are building something similar. That middle part – all your data, all emails, DMs, skills, know-how – is the company brain.

Humans are located at the edge of this brain, responsible for interacting with the real world. That is, humans are the point where this intelligent system touches reality. Humans can enter scenarios that models might not be able to handle yet. For example, meeting rooms, or novel, complex situations. I was going to use phone calls as an example, but now AI can easily enter phone call scenarios too.

More typical are unfamiliar situations, ethical judgments, and high-risk moments. For instance, a founder comes to us saying they are considering separating from their co-founder. In these truly high-risk, emotionally charged moments, you still want a human present.

That is the position of humans. For many of your companies, sales conversations are similar. For the next 20 years, I believe a human will still be needed in the room for sales.

So, I think humans will live on the edge of the company brain, responsible for bringing intelligence into the real world.

I've run out of time; the host is probably about to pull me off stage. I'll leave you with one final question: If you were to start your own company again today, would you design it this way from the start?

Most of you have companies small enough to do this perfectly well. So I think you have no excuse. And I know there are a few people here who are already tearing down and rebuilding their companies.

I'll stop here and hand it over to Pete. Thank you.

Video Link