The most interesting thing about Garry Tan writing code at 2 AM is the choice behind the habit. Founders have always had strange hours. The revealing part is what the head of Y Combinator chooses to build when his day job already puts him at the center of the startup world.

He is building a new operating model for software creation. The visible artifact is GStack, an open workflow for coding agents that turns tools such as Claude Code and Codex into something closer to a small software team. The deeper artifact is a philosophy: a founder who can describe a product clearly, test it quickly, and steer agents with good judgment can now move with the force of a much larger engineering group.

That is why the late night coding matters. Tan is stress testing the claim he keeps making to founders: AI has compressed the distance between idea and shipped product. If ten people can now do work that once required fifty, the strongest founders will be the ones who learn to manage agentic systems as carefully as they once managed human teams.

GStack makes that idea concrete. It adds structured roles around the coding agent: product review, engineering review, design review, code review, browser QA, release discipline, and retrospectives. In plain terms, it gives the agent a delivery loop. The human defines the problem, asks for critique, lets the agent implement, checks the result in a browser, then tightens the process for the next run.

This is a small but important shift. Many people use AI coding tools as faster autocomplete. Tan is treating them as junior organizations that need standards, memory, review, and taste. The output is code, but the real product is a repeatable way to turn intent into working software.

That same lesson applies far beyond web apps. A founder building an education tool might use Miss Formula to convert handwritten equations into clean digital formulas, then use ChatGPT to shape lesson explanations, and use Gemini to reason across multimodal materials. The magic comes from orchestrating specialized tools into a system that shortens the path from raw input to useful product.

This also explains why YC cares so much. The old startup filter rewarded credential, network, and the ability to recruit. Those still matter, but AI has made shipping speed more visible. A tiny team that pushes real product every day now creates stronger evidence than a beautiful pitch deck. Code commits, customer loops, product demos, and agent assisted iteration all become proof of momentum.

There is a warning inside the excitement. Agentic coding can create a lot of code very quickly, while good products require judgment, review, tests, security checks, and user taste. Without that discipline, speed turns into cleanup debt. That is why GStack is more interesting than a collection of prompts. Its value comes from the insistence that fast work still needs a process.

The image of Garry Tan coding at 2 AM is powerful because it captures the new founder posture. The best builders are becoming editors of machines, designers of workflows, and auditors of output. They write less glue by hand, but they make more decisions about what should exist, how it should behave, and whether the result is good enough for users.

So what is he building? He is building software, yes. More importantly, he is building a playbook for the AI native startup. In that playbook, the founder becomes the person who can point a swarm of capable tools at a real customer problem and keep raising the standard until something valuable ships.