Video Title: How to Build a Self-Improving Company with AI
Video Author: YC Root Access
Compilation: 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 deeper question: when AI is no longer just Copilot, but can perceive, decide, call tools, accept feedback, and self-correct, what should the company itself be redesigned into?

Tom's core judgment is that traditional companies still operate like "Roman legions": information relies on hierarchical upward communication, commands are distributed downward through management chains. But AI is breaking this organizational assumption. The truly important thing is not having engineers write 20% more code, but extracting scattered business knowledge from emails, Slack, meetings, documents, and human brains, turning it into organizational context that AI can read, call, and iterate on.

In his view, future AI-native companies will consist of a series of recursive, self-improving AI loops: systems perceive external changes from customer emails, support tickets, and product data, then make decisions through rule layers, tool layers, and quality gates, and finally learn and correct automatically based on results. YC is already experimenting with similar mechanisms: agents not only answer questions but also monitor which queries fail, determine if new tools, databases, or indexes are needed, and automatically submit code, review, merge, and deploy. In other words, the company can continue optimizing while founders sleep.

This also means that AI's impact on companies will go beyond tools, further transforming organizational structure. Tom proposes "burn tokens, not headcount"—the bottleneck for future startups may no longer be the number of employees, but token usage, quality of business context, and organizational knowledge readability. Middle management's coordination functions will be largely replaced by AI, while ICs, direct managers, and humans capable of high-risk judgment in the real world will become even more important.

Most notably, it's not just that AI makes companies more efficient, but that it is changing the very organizational form of "company" itself. When software can generate on the fly, processes can be automatically improved, and experiences can be continuously accumulated into the company's brain, the founder's goal may no longer be a clearly hierarchical team, but a set of intelligent systems capable of continuous learning and self-optimization.

Below is the original text:

Reimagining Operations: Companies Should No Longer Operate Like Roman Legions

This part is somewhat based on a previous talk by Diana. The weekend video is already online and very insightful. Also, Jack Dorsey posted some tweets about two or three weeks ago, which I found very interesting, so I "stole" many of his ideas and included them in this share.

This talk is more conceptual and high-level, mainly discussing how we should rethink building companies.

The design of the Roman legion was fundamentally about projecting power outward from Rome, covering two continents, and even extending to Hadrian's Wall near Scotland. It relied on a nested hierarchical structure, with each layer having a stable management span. Each level had clear leaders responsible for passing down commands and reporting information upward.

If you observe most companies today, you'll find they still operate like a Roman legion: people are the channels through which information flows up and down. One point that impressed me from Jack Dorsey's tweets is that we have always assumed hierarchical organizations are the best units for organizational economic value. But I believe AI is fundamentally breaking this assumption.

A year ago, if you asked people what AI is good for, they would usually talk about "productivity": for example, Copilot increasing engineers' efficiency by 20%, integrating Copilot into workflows to help teams deliver more software. But I think this is a flawed understanding. It’s like putting a more powerful engine into an old way of working. The real question is not how to add AI tools to old organizations, but to reimagine what a company is and how it should operate.

For example, what Garry just mentioned—I truly believe he can produce more code alone than an entire engineering team. What has been on my mind is how to extract internal domain knowledge and define it as context, skill sets, or whatever you want to call it.

Domain knowledge, business know-how, know-how—these are originally scattered in human brains, Slack messages, emails, Notion documents. These pieces of information collectively define how your company operates. Once you can make this knowledge clear and readable, you can shift from hierarchical organizations to a kind of AI-native, software-driven intelligent organization.

Making the company better while sleeping: How AI closed loops can automatically discover, fix, and deploy

AI is not an add-on outside the company. It’s not just a tool for engineers to boost efficiency. I believe we can reimagine the company as a set of recursive, self-improving AI loops. This is very important because once a company reaches this stage, it can even continuously optimize itself while you sleep.

Let me give an example.

Diana also mentioned this AI loop in her speech. It starts with a "sensor layer." This sounds fancy but is actually simple: customer emails, support tickets, code changes, unsubscribe requests, product telemetry data—these are all sensor data used to gather information from the external world.

Next is the strategy or decision layer, which is the rules: what AI can do, what requires human permission, what must be recorded. Then comes the tool layer, similar to Garry’s mention of skills and code, essentially deterministic APIs like database queries, calendar checks, and so on—tools that AI can call.

Then there are quality gates, such as Eva’s mention of deterministic checks, security filters, and human review for high-risk matters. Finally, there’s the learning mechanism: the system interacts with the real world, discovers where it’s not working, and feeds feedback back to the start of the loop.

If each step can run with little to no human intervention, then the system will keep improving even while you sleep.

Let me share some real examples from our current operations. Initially, we built an agent that you can ask questions, which has some deterministic tools to query our databases. For example, a simple question: when was the last time I held office hours with this company?

Later, it became smarter. For instance, I’m currently doing office hours with a company, and they need to connect with people in the petrochemical industry. The system can query databases in different ways, combine RAG methods, and find five relevant founders to recommend for networking.

But this is still just a sidekick, an assistant agent. It’s still the old way of using AI: making me more efficient as a group partner, boosting my work efficiency by 20% or 30%.

What really gave me an "aha moment" was adding a monitoring agent on top of this system. It checks every query initiated by YC staff, judges which succeed and which fail. Then it asks: why did it fail? How can we make this query succeed? Do we need new deterministic tools? Update skills files? Add new databases? New indexes?

These things now really happen automatically at night. It writes code, submits merge requests to YC’s codebase, which another agent reviews, then merges and deploys. So the next day, when someone asks the same question, the query succeeds.

For me, that’s the key moment. It’s not just making a human 20% or 30% more valuable. It’s AI completing the loop itself, finding ways to self-improve.

I believe that if you can identify parts of your company that can operate this way, and minimize human oversight and intervention, then you can invest tokens into this process, and the company will keep improving itself.

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

This cycle repeats over and over. Your product will have a self-optimizing loop.

Customer support is similar. Customer suggestions keep coming in, and you can use an agent to triage them. This agent acts like your chief product officer and CTO, making judgments: discard suggestions we don’t want to implement; prioritize those aligned with our roadmap, and deploy code to deliver to customers—all without human intervention.

So, if you can see every part of your company as a recursive AI self-improvement loop, it will be a completely different entity from a "Roman legion"-style hierarchy.

Less headcount, more burning tokens: AI-native companies will reshape organizational structures

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

First: spend tokens, not headcount. We see that many companies by Demo Day have already increased per-person revenue about fivefold compared to 18 months ago. I believe this trend will continue into Series A and B stages. Soon, your real constraint won’t be employee numbers but token usage.

The roughest approach now is to measure each person’s token usage. Of course, this metric can be gamed and is silly in extreme cases. But I think it’s the right direction. We are in an exploratory phase of "what is possible," so everyone should experiment as much as possible to see what this crazy new intelligence can do.

Once you turn it into a leaderboard and tie promotions or dismissals to this metric, it will be gamed and distorted. But in terms of direction, understanding who in your organization is pushing tokens to the limit and who isn’t is a way to decide where to focus your time.

I believe middle management is over. At least for coordination issues, I don’t think middle managers are needed anymore; AI should handle that.

For me, there are two key roles in the future. Jack Dorsey listed three, but I don’t like the third, so I removed it. I think only two roles matter: everyone must become an IC, a contributor, builder, operator. And crucially, there must be a directly responsible individual. Any initiative needs a clear person responsible, not a committee or a group.

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

By the way, I think most people are still at the forefront of this. I’d love to hear your progress. It feels like everyone is still exploring the boundaries. I’m not sure if anyone has already built truly self-improving companies in every function. Maybe I’m wrong, and you can prove me wrong.

If it were me, what would I do first?

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

Simply put, all emails from our partners, if you email a YC partner, that email goes into YC’s database. Every Slack message, DM, office hours—over the past three or four months, we’ve started recording everything. Everything that happens, as long as it’s recorded, is "happening" for AI; if not recorded, it didn’t happen for your intelligent system.

Just now, I was chatting with some founders, and we discussed many good things about their companies. Every time I think, I really should record this conversation. Because someone just asked me to introduce them to someone, and now I can’t even remember who that was. I promised to help, told them to email me later, but I know I’ll forget—I’ll be talking to 20 more people.

So, this might require phones, recording devices, smart glasses, or microphones in every room. Basically, everything needs to be recorded so AI can understand it.

Then, as Garry said, you need speaker separation and summarization. You can’t just dump 100k hours of recordings into a context window. You need to organize, aggregate, compress, and extract key parts, leaving some clues for AI.

For example: has anyone read YC’s user manual? I hope everyone in this room has at least opened it once. It’s okay if not. Most of that manual was written five to ten years ago and is somewhat outdated.

Last weekend, Harsh suddenly thought: since we’ve accumulated about 2,000 hours of office hours recordings over the past three months, why not generate a new version of the manual?

So, you can give the system a set of instructions: organize, compress, synthesize the recordings, categorize by topics like fundraising, recruiting, co-founder disputes, then have it write a new manual. By the weekend, it will produce a 150-page manual, clearly better than the current version.

Even more importantly, we can update it monthly. So, our manual becomes a self-improving system. Every new suggestion is compared with the existing manual—either integrated or discarded. This way, the manual becomes a living, continuously updated brain, carrying our weekly advice to founders.

Of course, it’s not just about the manual. You can feed it as context into an AI agent. Then you can ask a super-intelligent AI questions and get insights from the combined wisdom of 16 YC partners. But these insights must be readable by AI, so you must record everything.

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

The third is that every function should be able to generate its own software. In the past, we might call it a "dashboard," but now it’s about generating on demand. Codex 5.5 is good enough; most simple internal tools and dashboards can be generated at a high quality in one go. I tested some of our internal tools over the weekend, and the results were incredible.

So, all internal operations teams should sit on top of this layer: have an intelligent understanding of the business, then generate dashboards and workflows themselves.

And I see these softwares as disposable. The truly valuable thing is data. As Garry said, he stores all emails as Markdown, never discarding anything. But software itself is ephemeral and temporary—you can generate it, and you can regenerate it.

What’s truly valuable is the human understanding of the business: how a function operates, how we run YC events, etc. The software used to execute these activities can be generated for each event and discarded afterward. After a month or two, the model gets smarter, and you can throw away the old software, give it the original instructions again, and regenerate new software.

So, I believe that business context and skills are valuable. The software built on them is transient.

In this world, what is the role of humans?

I think we’re talking about a "company brain." I know many of you are doing similar things. The middle part—your data, emails, DMs, skills, know-how—is the company brain.

Humans are on the edge of this brain, responsible for interacting with the real world. That is, humans are the point of contact between this intelligent system and reality. Humans can enter scenarios that the model can’t yet handle, like meetings or complex, novel situations. I initially thought of phone calls as an example, but now AI can also easily handle phone scenarios.

More typical are unfamiliar situations, ethical judgments, high-risk moments. For example, a founder comes to us considering splitting from a co-founder. In these high-stakes, emotionally charged moments, you still want a human present.

That’s where humans fit in. For many companies, sales conversations are similar. I believe that in the next 20 years, a human will still be needed in the room during sales.

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

I’ve gone over time, and the host might be about to cut me off. Finally, I leave you with a question: if you were to start your own company today, would you design it from the beginning as this kind of system?

Most of your companies are still small enough to do so. So I see no excuse. And I know some of you are already tearing down and rebuilding your companies.

That’s all I’ll say for now. I’ll hand it over to Pete. Thank you all.

[Video link]

Click to learn about Rhythm BlockBeats job openings

Join the Rhythm BlockBeats official community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Group: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia