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 partner Tom Blomfield discusses not "how to use AI to improve employee productivity," but a more fundamental question: What should a company be redesigned to look like when AI is no longer just a Copilot, but can perceive, make decisions, call tools, accept feedback, and self-correct?
Tom's core judgment is that traditional companies still operate like "Roman legions": information is transmitted upwards through hierarchical layers, and commands are distributed 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 human minds, and turning it into organizational context that AI can read, call, and iterate.
In his view, future AI-native companies will consist of a series of recursive, self-improvement AI loops: systems will perceive external changes from customer emails, customer service tickets, and product data, then execute decisions through rule layers, tool layers, and quality checkpoints, and finally learn and adjust automatically based on results. YC is already experimenting with similar mechanisms: agents not only answer questions, but also monitor which queries fail, determine whether new tools, databases, or indexes are needed, and automatically submit code, review, merge, and deploy. In other words, the company can continue to optimize while the founders sleep.
This also means that the impact of AI on companies will not be limited to the tool layer, but 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 the use of tokens, the quality of business context, and the readability of organizational knowledge. The coordination functions performed by middle management will be greatly replaced by AI, while ICs, direct responsible individuals, and human roles capable of making high-risk judgments with the real world will become more important.
The most noteworthy change is not that AI makes companies more efficient, but that it is changing the very form of the organization called "company." When software can be generated on-demand, processes can be improved automatically, and experiences can be continuously distilled into the company brain, what founders really need to build may no longer be a clearly hierarchical team, but an intelligent system capable of continuous learning and self-optimization.
The following is the original text:
Rewriting Operations: Companies Should No Longer Operate Like Roman Legions
This section is somewhat based on a previous talk by Diana. The video from the weekend is now online and is very exciting. Additionally, about two to three weeks ago, Jack Dorsey posted some tweets that I found interesting, so I've "stolen" quite a few of his ideas and incorporated them into this sharing.
This sharing is relatively conceptual and high-level, mainly discussing how we should rethink company building.
The design of the Roman legion was essentially to project power outwards from the center of Rome, covering two continents, even extending to areas like Hadrian's Wall near Scotland. It relied on a nested hierarchy structure, where each layer had a stable management span. Each level had a clearly designated responsible person, who was responsible for passing commands downwards and relaying information upwards.
If you observe most companies today, you'll find that they still operate like a Roman legion: people are the channels through which information flows. One point that struck me from Jack Dorsey's tweets is that we have always assumed that hierarchical organizations are the best way to organize economic value units. But I believe that AI is fundamentally breaking this assumption.
A year ago, if you asked people what AI was useful for, they would typically talk about "productivity": for example, Copilot improving engineer efficiency by 20%, integrating Copilot into workflows to help teams deliver more software. But I feel this is a problematic way of understanding it. It's like putting a more powerful engine on an old way of working. The real question is not how to add an AI tool to an old organization, but to reimagine what a company is and how it should operate.
For example, what Garry just mentioned, I truly believe that one person can now produce more code than an entire engineering team. What I keep thinking about is how to extract domain knowledge from within the company and define it as context, skill sets, or whatever you wish to call it.
Domain knowledge, business knowledge, know-how, are originally scattered in human minds, Slack messages, emails, Notion documents. This information collectively defines how your company operates. Once you make this knowledge clear and readable, you can shift from a hierarchical organization to an intelligent organization driven by AI-native software.
Let the Company Improve While You Sleep: How AI Loops Automatically Discover, Repair, and Deploy
AI is not an add-on thing next to the company. It is not just a tool for engineers to use to improve efficiency. I believe we can reimagine the company as a set of recursive, self-improving AI loops. This point is very important because once a company reaches this stage, it will continue to self-optimize even while you sleep.
For example.
Diana mentioned this AI loop in her talk. It starts with a "sensor layer." This term sounds advanced, but it can be quite simple: customer emails, customer service tickets, code changes, user cancellations, product telemetry data—these are all sensor data used to gather information from the external world.
Then there is the strategy or decision layer, which consists of rules: what AI can do, which actions must request human permission, and which actions must be recorded. Further down is the tool layer, akin to the skills and code mentioned by Garry, essentially deterministic APIs, such as querying databases, checking calendars, etc., which is a set of tools that AI can call upon.
Next is the quality checkpoints, such as deterministic checks, safety filters, and human reviews of high-risk matters. Finally, there is the learning mechanism: the system interacts with the real world, discovers where it does not work, and then feeds the feedback back to the start of the loop.
If each step can operate without human intervention, or with minimal human intervention, then the system will become better while you sleep.
I can give you some examples of what we are currently running. Initially, we created an agent that you can ask questions; it has some deterministic tools to query our database. For example, a very simple question: when was the last time I had office hours with this company?
Lately, it got a bit smarter. For instance, I am having office hours with a company, and they need to meet people related to the petrochemical industry. This system can query the database in various ways, combine methods like RAG, find five relevant founders, and recommend them to you.
But this is still just a sidekick, an assistant agent. It is still last year's way of using AI: AI makes me 20% or 30% more efficient as a group partner.
What gave me the "aha moment" is when we added a monitoring agent on top of this system. It looks at every query initiated by each YC employee, determines which queries were successful, and which ones failed. Then it asks: why did it fail? How can this query succeed? Do we need a new deterministic tool? Do we need to update the skills document? Do we need a new database? Do we need a new index?
These things now really happen automatically at night. It writes code, submits merge requests to YC's code repository, lets another agent review, then merges and deploys. So the next day, when a human comes back to ask the same question, that query can succeed.
To me, this is the key moment. It doesn't just make a human 20% or 30% more valuable. Rather, AI completes the loop and finds a way to self-improve.
I believe that if you can identify which parts of the company can operate this way, and minimize human roles in execution and oversight as much as possible, then you can invest tokens into this issue, and the company itself will continue to improve.
There are many other examples. For instance, if you have product analytics data, you can have an agent analyze product data to identify the friction points in the sales funnel. It can study best practices, set up an A/B test, run it for a week, select the best-performing version, and then deploy it.
This will happen over and over again. Your product will have a self-optimizing product loop.
Customer service is similar. As customer suggestions come in, you can use an agent to triage. This agent functions somewhat like your Chief Product Officer and Chief Technology Officer; it has to make decisions: we don't want to implement this suggestion, throw it out; but that suggestion fits with our roadmap and can be completed tonight. Then it writes code, deploys, and delivers directly to the customer, completely without human intervention.
So, if you can see every part of the company as a self-improving recursive AI loop, it will become a very different entity from a "Roman legion-style" hierarchical company.
Burn Tokens, Not Heads: AI Native Companies Will Reshape Organizational Structure
So what does it mean if you want to do this?
The first point is: burn tokens, not heads. We now see that many companies have achieved about five times the revenue per person by Demo Day compared to 18 months ago. I believe this trend will continue into Series A and Series B stages. Soon, you will find that your true limitation will not be the number of employees, but the amount of token usage.
Now the roughest practice is to measure each person's token usage. Of course, this metric can be silly in extreme cases and is easily gamified. But in terms of direction, I believe it is the right one. We are currently in a phase of exploring "what is possible," so everyone should experiment to the fullest and see what this crazy new intelligence can do.
Once you make it a leaderboard and tie promotions or dismissals to this metric, it will certainly be gamed and distorted. But in terms of direction, figuring out who is maximizing token usage in the organization and who isn't is indeed a way to determine where you should focus your time among employees.
I think mid-level management is over. At least in terms of coordination issues, I don't believe mid-level management is still needed; AI should take care of this.
To me, there are two important roles in the future. Jack Dorsey proposed three, but I don't really like the third one, so I deleted it. I believe the truly important roles are: everyone must be an IC, meaning individual contributor, builder, operator. And the key is to have a directly responsible individual. For anything to advance, there must be one clearly designated person in charge, not a committee or a group of people.
I believe a company can completely be built on ICs. Mid-level management is truly over. And building a self-improving company is this vision.
By the way, I feel that everyone is still at the forefront of this matter. I am also very curious to know how far along you all are. It feels like everyone is still exploring the boundaries. I'm not sure if anyone has established a truly self-improving company in every function yet. Maybe I am wrong, and you can prove me wrong.
What would I do first?
The very first important thing is to make the entire organization readable and understandable to AI. What does this mean? It means you must document everything.
In simple terms, now all partner emails, if you email a YC partner, that email will go into the YC database. Every Slack message, every DM, every office hours session—we have started recording all of this over the past three to four months. Everything that happens, as long as it is recorded, happened for AI; if it is not recorded, it hasn't happened for your intelligent system.
Earlier, I was chatting with some founders here, and we discussed a lot of good content about their companies. Every time I chat, I think, I really should record this conversation. Because someone needed me to introduce them to someone, and now I can't even remember who that introduction was for. I promised him I could help and then told him to email me later because I knew I would forget; I have to chat with 20 more people.
So, this may require using phones, recording devices, smart glasses, or equipping every room with microphones. In short, everything needs to be recorded so that AI can comprehend it.
Then, as Garry said, speaker separation and summarization need to be done. You cannot just stuff 100,000 hours of recordings directly into the context window. You must organize, aggregate, compress, and distill them into important parts, and then leave some clues for AI.
For example: has anyone here read the YC user manual? I hope everyone in this room has at least opened it once. That's okay. 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 regenerate a version of the user manual?
So you can give the system a set of instructions to first organize, compress, and aggregate the recordings, then categorize them by themes such as fundraising, recruitment, founder disputes, etc., and then let 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 clearly of better quality than the existing version.
More importantly, now we can update it every month. Thus, our user manual becomes a self-improving system. Every new suggestion will be compared with the existing user manual, either being absorbed into it or discarded. In this way, the user manual becomes a continuously updating living brain, carrying the advice we give to founders every week.
Of course, it will not remain confined to just the user manual layer. You can use it as context input for AI agents. Suddenly, you can ask a super-intelligent AI questions and get a synthesis of wisdom from 16 YC partners. But the prerequisite is that this knowledge must be readable by AI. Therefore, you must document everything.
The second point is similar: if something can create a self-improving artifact that can be read by AI, then 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 talked about "dashboards," but now it is not just dashboards but on-demand generated software. Codex 5.5 is now good enough that for most simple internal software and dashboards, you can generate them at a high quality in one go. I tried some of our internal stuff over the weekend, and the results were truly amazing.
So, all internal operations teams should sit above this layer: having a smart understanding of the business, and then generating dashboards and workflows themselves.
Moreover, I would treat this software as completely disposable. What should be preserved with great care is the data. Just as Garry said, he saves all emails as Markdown and never discards anything. But the software itself is temporary and ephemeral. You can generate it, and you can regenerate it.
What is truly valuable is the understanding of the business in human brains: how this function operates, how we conduct a YC event, and so on. As for the software used to execute events, you can generate one for the event, use it, and then discard it. A month or two later, when the model has become smarter, you throw away the old software, give it the original instructions again, and generate a new version of the software.
So I believe that what is valuable is the business context and skills. The software built upon them is temporary.
So what is the role of humans in this world?
I believe what we are discussing is a "company brain." I know there are many people in this room doing similar things. The middle part—your data, all emails, DMs, skills, know-how—is the company brain.
Humans are situated at the edge of this brain, responsible for interacting with the real world. That is to say, humans are the point of contact for this intelligent system with reality. Humans can enter scenes that the model cannot temporarily access. For example, in a meeting room or some novel, complex situations. I initially intended to use telephone as an example, but now AI can easily enter telephone scenarios too.
More typically, it involves unfamiliar situations, ethical judgments, and high-risk moments. For instance, if a founder comes to us and says they are considering parting ways with their co-founder. In these genuinely high-risk, high-emotion situations, you would still want a human present.
This is the position of humans. For many of your companies, sales dialogues are the same. I believe that in the next 20 years, there will still be a need for a human in the room during sales interactions.
Therefore, I believe humans will live at the edge of the company brain, responsible for bringing intelligence into the real world.
I have exceeded my time, and the host may soon pull me off the stage. Finally, I leave you with a question: If you were to start your own company today, would you design it to be in this form from the very beginning?
Most of your companies are still small enough to do this. So I feel you have no excuse. And I know several of you in the audience are in the process of dismantling and rebuilding your companies.
So I'll wrap it up here and hand the time over to Pete. Thank you all.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
