The latest sharing from YC's leader: In the AI era, never do one-off work
In 2026, as generative AI dives deep into real-world applications, the global entrepreneurial ecosystem is undergoing a fundamental restructuring. The traditional business logic of driving output by piling up headcount is becoming obsolete, replaced by an entirely new technology-native paradigm.
Garry Tan, the head of Y Combinator, pointed out in his latest speech that the true dividend of the AI era is by no means fine-tuning model weights, but the complete reengineering of workflows. By fully encoding employees, processes, and performance within traditional organizational structures into "Skill Manuals" and "Company Brain", Silicon Valley has already seen the first batch of AI-native enterprises with extremely lean teams of 15 or 40 people generating tens of millions to hundreds of millions of dollars in ARR (Annual Recurring Revenue), completely overturning the per-capita output rule of traditional business since the Industrial Revolution.
"If you still treat AI as nothing more than auto code-completion and prompts as one-off tools, you will get your lunch eaten by competitors in the next phase of the race." Garry Tan remarked.
Key core takeaways from the speech:
What determines the upper limit of productivity in the AI era is not the weights of large models themselves. The real performance leverage — which can deliver up to 400x improvements — lies in how you reconnect and reweave your workflows.
The traditional business logic of driving output by piling up headcount is failing. Future organizational forms will be fully codified: one "Skill Manual" equals one virtual employee, and a "Parser Table" becomes your organizational chart.
In system development, you must strictly clarify the boundaries of computation: delegate non-deterministic tasks such as intuitive judgment to the "Latent Space" of large models, while leaving explicit rules and code to the "Deterministic Space", to avoid misalignment between the two.
You must break through the limitations of human "working memory", and build a "Company Brain" through context engineering, precipitating the enterprise's historical knowledge into a "Knowledge Library" that can be precisely retrieved and generates compound value over time.
Never do one-off work. Once you are satisfied with the AI's output, you must "skillify" it into a reusable process, otherwise your enterprise will fall into a state of "technical amnesia" every single day.
We have compiled the key highlights of this speech, here is the full transcript:
1. 400x Performance Leap: From Individual Productivity to Workflow Reconstruction
Garry Tan: Good afternoon everyone, how are you all doing? Ready for this transformation? Theo just asked a very critical question: What should we be building right now? As a founder, an investor, and the leader of a 20-year-old institution that is transitioning to become AI-native, I want to answer this question from an investor's perspective. Leading a 20-year-old organization to become AI-native is a peculiar yet wonderful experience. Over the next roughly 20 minutes, I will talk about what YC is trying to do. We want to nurture enterprises where one single person can accomplish work that used to require two people, or even a thousand people to complete. This is not a metaphor. In practical terms, everyone in this room can achieve this this year. In about an hour, some of you will step onto the entrepreneurial battlefield, and I hope you walk in with a clear understanding of what today's technology is capable of — because the potential of current technology is far greater than most people imagine.
Garry Tan: Let me start with a number. Although I've faced widespread online criticism for this viewpoint, I will restate it in front of all of you today. This room is the best place in the world to stress-test this idea, and I would much rather verify it in person with all of you. Back in 2013, when I was a YC partner, I was mainly responsible for building YC's internal social network. At that time, while I did make investments, I was essentially a full-time engineer. Back then, I could only write about 14 lines of truly useful logic code per day — that number excludes all comments and blank lines. If you look at technical literature from that era, you will find this was completely normal: some people wrote 15 lines, others 50 lines, nothing like the thousands of lines of code many young developers today produce on a daily basis. Those 15 lines represented my median daily output even when I was operating at full capacity. This year, I run YC full-time. It's still me, putting in roughly the same amount of time — in fact, I work even fewer hours because I have to pick up my kids at 5 PM — but when I recalculated my total output, I found that my efficiency has increased by approximately 400 times.
Garry Tan: Before the skeptics jump in to question this number, let me break it down for you myself. If you don't trust the pure line-of-code metrics, that's fine. You can apply the strictest redundancy penalty mechanisms to strip out all fluff: assume the AI Agent produces bloated code, assume half of it is just boilerplate, even assume I'm tooting my own horn — at the very baseline, you still get an 8x improvement, and the median is 80x. No matter how you slice and examine it, this is an enormous figure. Here is a critical point, something I would almost want to brand onto the inside of everyone's eyelids if I could: This difference does not come from the models themselves. People who deliver 2x performance and those who deliver 100x performance are using the exact same Claude model, identical weights, the same context window, and the same API. Therefore, the leverage does not lie in the model weights — it lies in how you connect and arrange your workflows.
Garry Tan: And this is not something that only happened to me. At YC, we see this all the time. In the Winter 2025 batch, one quarter of the companies had 95% of their codebase generated by AI — and that was last year. That cohort has already become YC's fastest-growing, most profitable batch in history. Across YC's history, 94 companies have successfully broken $100 million in revenue after a single seed round. So I think we know what we are talking about. While I cannot definitively prove that AI-generated code directly caused this growth, I can tell you that the fastest-growing founders we backed did not treat AI merely as an auto code-completion tool — they treated it as a productive workforce. Those companies that re-wove their workflows completely are the ones that truly pulled ahead of the curve.
2. The New Physics of Labor: Codify Your Organization into "Skill Manuals"
Garry Tan: So what does it actually mean to rebuild your workflows? This is the core of this speech, and the part I most want you to take away. Everything we learned from building AI Agents can be directly mapped one-to-one onto real-world organizational structures.
Garry Tan: In the AI era, one Skill Manual is equivalent to one employee. It embodies one capability, responsible for one specific job, documented so clearly that any entity can execute it. Meanwhile, the Parser Table is your organizational chart — when a task comes in, the parser decides who handles it and where to route it. Archiving Rules are your internal business processes, used to verify that the parser is working properly and remaining compliant. Trigger Evaluations are your performance reviews. Every single functional module of an organization that once required thousands of people to run can now be represented as code files and other types of descriptive documents. We have been building organizations all along — we just previously lacked this kind of management layer, and now we have it. When you sit in front of your computer and collaborate with Claude, Coder, and Codex to develop, you are not just writing software. You are essentially recruiting, training, and managing a workforce made of code and documents.
Garry Tan: I have seen countless companies putting this into practice. For example, Emergence — an AI application building tool project from the Summer 2024 batch — took only eight months from public launch to reaching nine-digit annual recurring revenue (ARR). When their ARR exceeded $15 million, their team had just 15 people. By Winter 2024, their retail revenue hit $60 million with a team of roughly 40 people. This extremely high per-capita output was unimaginable in the past — it never existed in the software industry, the oil industry, or the railroad industry. They are not some kind of industry anomaly; they are simply the first batch of companies natively built on top of this new physical law. So how exactly do these companies operate? They do not hire hundreds of employees for departments like sales, support, operations, and finance. All the AI-native companies I have observed inside YC codify all these functions into skills and written processes. The only reason they hire engineers is to maintain these skills, and solve problems that AI Agents cannot yet handle. This is what a real AI-native company looks like — it is absolutely not a thought experiment. If you configure the corresponding Skill Manuals properly, it can even help you file your taxes.
Garry Tan: Now imagine the YC Startup School scene — it feels very similar to this room. 400 companies or 400 founders sitting at long tables, I can picture each of you busy working on a laptop every day. The amount of work you get done in one single day now equals what a person used to accomplish in an entire year. This is not some distant future — it is the immediate baseline the industry can reach right now. If you don't do this, your competitors will, and they will take your lunch very gracefully, and even politely thank you for it. This is the extension point that most engineering stacks easily overlook: this does not only affect engineers. During YC's digital transformation, our media staff, event employees, and finance teams are all participating. Employees who have never opened a terminal in their lives are now building their own Skill Manuals and scheduled tasks. One of our finance colleagues, using our internal open-source system and Company Brain, recently integrated around 100 complex Excel workbooks into an application she built herself. She is not a programmer, but she is now a manager overseeing AI Agents — in fact, everyone at YC is now in that position.
Garry Tan: This is why YC can maintain such a huge operating scale with a tiny team — a team size that would seem negligible to any other company of equivalent size. It is not because we work harder; it is because we have a completely different organizational form. This is the core of winning. This is not just 400x more efficient engineers — this is a company that operates at 400x the efficiency.
3. Engineering Duality: Clarify Computational Boundaries Between Latent Space and Deterministic Space
Garry Tan: If you can only remember one thing from this, it is that during development, you must be extremely, extremely careful to clarify exactly where computation is happening. Computation almost always occurs in two distinct spaces. All the AI engineering flaws and problems we encounter typically stem from content that was supposed to run in one space mistakenly ending up in the other. The first space I call Latent Space — which is the actual large language model itself. What do you use it for? The answer is non-deterministic calls for aesthetics, intuitive judgment, and understanding the real intent behind people's ambiguous statements. These computations reside inside the model, and you guide them through descriptive documents. The Deterministic Space, on the other hand, is a domain that engineers are very familiar with: for example, your code Agent will automatically write TypeScript, or if you are using Elixir, they will write Erlang code. The Deterministic Space is the second place where computation happens.
Garry Tan: Let me give you an example — this is a real pain point we are facing in our upcoming Startup School. We have 6,000 attendees, and one experiment we are trying to run is: can we seat 800 people simultaneously and achieve perfect clustering? That means the people sitting to your left and right are exactly the perfect contacts you should meet at the startup school. To make this happen, we must combine the Deterministic Space with the Latent Space. This computation and specific data storage — like where exactly each person sits in this 800-seat multi-dimensional array — absolutely cannot be stuffed into a context window. The large language model only needs to handle the perceptual part: arranging people into seats. This follows exactly the same logic a human would use to perform this task. If a human had to do this, they would probably print out 800 pages of materials, spread them out across a big room, and figure out where each person should go. But now all of this work can be done on a computer. A task that would have taken a month to complete can now be done in 10 minutes with a few hundred dollars worth of Tokens. I think this is absolutely remarkable — something that would have been impossible even half a year ago.
4. Defining the Company Brain: A "Library and Librarian" That Breaks the 7-Digit Working Memory Limit
Garry Tan: This brings us to the concept of working memory, which is my personal favorite way to understand it. The human brain can only process about 7 things at a time — the famous 7±2 rule. This is one of the most well-known papers in cognitive psychology, and it explains why local phone numbers are usually 7 digits, and why you always forget the 8th item on your grocery list. This is basically the full extent of an individual human's working memory. Every institution, every checklist, every organizational chart, every filing cabinet that humans have ever built throughout history is essentially a prosthetic invented to compensate for this limitation. The logic behind this is actually pretty crazy when you think about it carefully. But an AI Agent can hold 1 million Tokens — roughly equivalent to 1,000 pages of content. I recently tried to explain what a Company Brain is to my 10-year-old kid: I told them an AI Agent can have three copies of "Harry Potter" open in its mind at the same time, find any detail across all of them in seconds, and perform a comprehensive analysis of the content of all three books. That is truly amazing. Three "Harry Potter" books against the human 7-digit limit — that is mind-blowing. Is this AGI? Maybe not quite, but it already demonstrates a completely different operating and governance mechanism. However, almost every single company on Earth today is still operating with traditional organizational structures designed for 7-digit human brains.
Garry Tan: But this also reveals a fact to us: three books is a lot, but it's actually very little. Your company does not have just three books — your company is a library. It contains every email, every meeting, every decision and its derivation process, every customer conversation, every retrospective report. The difference between whether your Agent is a genius or a goldfish with a 7-second memory comes down to who decides which three books to open on the desk. That is Context Engineering. And this is what we call the Company Brain: it is both the library and the librarian. Now some of you might be thinking, isn't this just Retrieval-Augmented Generation? You are right. Retrieval is the most fundamental underlying layer, just like how the bottom layer of Postgres is nothing but B-trees. But the really hard part is everything surrounding it: what content initially gets deposited into the wiki, how that content gets enriched and linked, what content gets promoted to hot memory, what gets archived as cold reference data, and who arbitrates when two conflicting facts emerge. Retrieval itself is not difficult — the hard part is building yourself a product that is actually worth retrieving from.
Garry Tan: I have been publicly building my own system called G-Brain. It works with any development framework, but it fits best with Open Claw and Hermes Agent. It is essentially Postgres for Agents — a dedicated retrieval layer whose job is to calculate exactly which three books should be loaded into the Agent's brain for any given task. My personal system started with a few rooms of books, and now it has grown into a warehouse containing roughly 220,000 pages of content. This content is mostly automatically written by my Agent based on my emails, meeting notes, 20 years of personal notes, and my life experiences. That's the point: it has become my second brain. When a founder emails me about a crisis,