In this video, Y Combinator CEO Garry Tan and General Partner Diana Hu explain how the way companies are founded and run is changing in the AI era. They emphasize that productivity is improving dramatically through the combination of AI agents and code, opening up new opportunities for founding companies. Just as YC's SAFE agreement once set the standard for venture capital, they argue—through that analogy—that AI agents are now creating the new standard, underscoring the revolutionary change AI will bring.

1. What Makes the Stanford CS153 Lecture Special, and YC's Influence

The Stanford CS153 lecture is more than a class—it is a special venue that distills the teachings of Silicon Valley leaders. This lecture carries on the spirit of Silicon Valley, like Peter Thiel's "Zero to One" or Sam Altman's startup lecture before it. It is especially meaningful that Garry Tan is once again sharing the experience he gained as a YC partner, when he had a major influence on the startup ecosystem.

Drawing on his experience as a Stanford student, Garry Tan spread Stanford's spirit through Silicon Valley, and now, together with Diana Hu, he is updating YC's new philosophy. As the host put it, it's "a closing-the-loop kind of moment"—his return shows the organic relationship between Stanford and Silicon Valley.

2. SAFE, the Pioneer of Capital Standardization

The lecture begins by emphasizing that system design can be applied to every field, not just engineering. Just as the standardization of electricity once led the Industrial Revolution, Silicon Valley needed standardization in capital allocation. When Garry Tan came to Silicon Valley in 2011, the venture capital market was sheer chaos.

"When I came to Silicon Valley in 2011, venture capital was in the era before capital standardization. It was a total mess."

It was then that Paul Graham and Jessica Livingston introduced a new standard called SAFE (Simple Agreement for Future Equity). At the time it looked like a simple legal document, but in the end SAFE became the standard for early-stage startup fundraising, and YC established itself as the institution that standardized seed-stage funding. It's said that without SAFE, Silicon Valley's development would have been very different from today.

3. The New Standard of the AI Era: Agentic Coding 🤖

The core message of the lecture is that AI is creating a new standard. If the standard of the past was a legal document, now code is taking over that role.

"SAFE was a legal instrument. What we're going to talk about today is code. And not just code—Markdown is code too."

Garry Tan emphasizes that with the advent of AI, the unit of production is changing. Where previously you needed fundraising and a lot of hiring, now even a small number of people can achieve enormous results.

"AI will change the unit of production. When I was sitting in your seat, I thought you had to raise money and hire a lot of people. But not anymore."

Citing his own experience as an example, he says that Posterous—a blogging platform he once built over two years with ten people and $4 million—he was recently able to rebuild in five days using the $200 Claude Max plan. Such productivity gains support Steve Yegge's claim that AI coding agents bring 10x to 100x, even 1000x, productivity improvements.

"I was able to build all the software in five days with the $200 Claude Code Max plan. Any one of you can do it—it doesn't take two years, just five days."

Garry Tan rebuts the criticism that AI merely produces "AI slop"—sloppy output—explaining that with a rigorous verification process such as securing 80–90% test coverage, you can produce high-quality code usable in real production. His Gstack project has received over 100,000 GitHub stars and is used by more than 15,000 people every day. 🤯

4. Core Elements of an Agentic System: Skills, Resolvers, Skillify, Evals, Memory 🧠

Garry Tan introduces the core concepts needed to build an agentic system.

4.1. Skill

A skill means extracting a specific persona from an LLM to do work. Just as YC's office-hours idea-validation process is summarized into the questions "What's the problem, who's the customer, how do you know, what are you building?", a skill contains the core knowledge and procedures needed to perform a specific task. It's likened to a runbook—something you jot down in a notebook each time you have to hold an event many times.

"A skill is basically a runbook. If you have to keep holding an event, what would you do? You'd write it down in a notebook. 'One, I need to secure a venue. Two, I need to figure out who should come.'"

4.2. Resolver

A resolver is a mechanism that dynamically loads the instructions or information an agent needs when performing a specific task. Like a master directory, the agent loads relevant information only when needed, reducing the burden on the context window. This is very important for running an agent system efficiently.

4.3. Skillify and Eval 🛠️

Skillify is the process of automatically turning a successfully completed task into a skill and testing it. This process goes beyond simply writing code and includes various verification steps.

• Unit Tests: Tests for the actual code.
• LLM Evals: Evaluation of the skill file.
• Integration Test: System integration testing.
• Resolver Trigger: Checking whether the agent fires.
• LLM as Judge Eval: Checking whether the trigger's breadth is appropriate.
• DRY (Don't Repeat Yourself): Preventing duplicate skills.
• End-to-End Smoke Test: Testing the entire system.
• Schema: Defining where memory is stored.

He explains that such a complex verification process is similar to why audit and compliance are needed in human organizations.

"When I was sitting in your seat, I didn't understand why so many people in so many human organizations had to spend so much time on audit and compliance. But now, at 45, building so many agentic systems and seeing how much time goes into skillify, I finally understand."

In particular, cross-modal eval uses several state-of-the-art models (Opus, GPT-5.5, DeepSeek-V4) to evaluate inputs and outputs, then feeds the results back to sub-agents to automatically iterate and improve. This becomes the core mechanism by which an agent system learns and evolves on its own.

4.4. G brain and a Three-Tier Memory System 🧠

Garry Tan describes the G brain he is developing as a three-layered memory system. This is a concept that evolved from the knowledge wiki Karpathy mentioned.

1. Knowledge Wiki: Basic information storage.
2. Vector Search, RRF Fusion, Backlinks: Strengthening information retrieval and connections.
3. Graph Database: Storing relational information through a knowledge graph.
4. Epistemology System: Tracking the source and reliability of information so that you can trace a particular person's intuition or belief and observe the process by which it actually gets implemented. This plays an important role in tracking how a founder's intuition becomes reality.

"I want my knowledge system to be able to track that 'someone believed X and made it real.'"

5. The AI-Native Company: A Closed-Loop System 🔄

Diana Hu explains that what fundamentally distinguishes an AI-native company from existing companies is that it runs as a closed-loop system. In existing companies, information is scattered in people's heads, informal conversations, meeting notes, and so on, making them inefficient like an open-loop system. When errors accumulate, the entire system can collapse.

But an AI-native company embeds agents into every decision-making process, creating a rigorous feedback loop. Agents have read access to every artifact the company produces, and through this they propose the next task to perform or bug fixes to make. This means the company becomes a system that heals and improves itself.

"One of the core ideas of building an AI-native company is that you have to fundamentally change the way the company operates. Generally, today's pre-AI companies basically run as open loops."

"Now we have the ability to turn all of this into a closed-loop system, integrating the agents and implementation approach Garry described into the company's decision-making."

Because of this change, companies are emerging that achieve astonishing results of $1–2 million in revenue per employee. Compared with existing companies, whose per-employee revenue is under six figures, this means a productivity gain of at least 10x.

5.1. A New Organizational Structure: Builder, DRI, AI Founder 🏢

In an AI-native organization, the role of middle managers shrinks, and three core roles become important.

1. Builder: Everyone becomes a builder who creates something. Even non-technical staff can build directly using tools.
2. DRI (Directly Responsible Individual): The person who bears direct responsibility for every outcome. They collaborate and coordinate with all team members to achieve the goal. Often the founder takes on this role.
3. AI Founder: The person who runs the company fast at the frontier of the future, leveraging every tool. They continuously absorb constantly changing technology and bring it into the company.

"Everyone becomes an Individual Contributor who builds something. Non-technical people now gain the power to build with all these tools."

"The DRI collaborates with ICs to coordinate so the goal gets achieved. For example, if the company's goal is 'triple revenue by the end of this week,' the DRI is responsible for coordinating everything needed to achieve it."

"A new role emerging in AI-native organizations is what we call the 'AI Founder.' Garry Tan is exactly this kind of person. You have to run the company fast at the frontier of the future, leveraging every tool."

5.2. The Key to Avoiding AI Slop: 'Taste' and Evals ✨

In building AI systems, how to avoid 'AI slop' comes down to 'taste' and rigorous evals. The cost of writing code will approach zero, but the taste to judge what is good and bad cannot be automated.

"What you need to build these agentic systems to avoid AI slop is the concept of 'taste.' You've heard a lot that 'taste' will endure, right?"

"The cost of shipping code will converge to zero, but the taste to make something good and to tell good from bad will not converge to zero."

Evals must go beyond mere general benchmarks. You have to accurately evaluate whether the product actually works, whether users are satisfied, whether it achieves business goals, and whether it complies with domain rules. For this, a human in the loop is essential—you have to detect when an error occurs and steer the system in the right direction.

6. The Golden Age of Founding: Hidden Opportunities 🚀

Diana Hu emphasizes that now is the best time in history to start a company. Among YC portfolio companies, firms like Salient (a voice agent for lending services), Happy Robot (a freight agent), and Reducto (document processing) are showing astonishing growth, going from zero to tens of millions of dollars in revenue in a single year. They are not merely building AI demos—they are deploying complete solutions and solving the market's painful problems.

In particular, there is vast white space to use AI in areas outside of computer science (CS), such as back office, finance, data, and customer service.

"In all these other areas outside computer science—back office, finance, data, academia, cybersecurity, customer service—there is enormous white space. There's room here for hundreds of AI unicorns to be born."

According to YC's statistics, in the past only the top 1% of companies recorded 10% weekly growth, but now many companies are achieving 3x growth in three months on average. This signifies the unprecedented growth opportunity that the AI era has brought.

Closing

This lecture shows that AI technology is fundamentally changing the paradigm of founding and running companies, beyond being a mere tool. The combination of AI agents and code, the closed-loop system, the new organizational structure, and the importance of human 'taste' and evals are the core elements for successful founding in the AI era. It concludes by delivering the hopeful message that this is truly the 'first inning of the revolution,' and that every Stanford student listening can stand at the vanguard of this revolution. 🌟