What can founders learn from Garry Tan?

His core lesson is to run AI coding agents the way a good manager runs an engineering team, with roles, process, and review, because out of the box the model wanders and guesses. He also models pressure-testing an idea before building it (his Office Hours skill forces the question of what evidence proves someone wants the product), staying a hands-on builder even as a CEO, automating the drudgery you personally feel, and acting on the fact that the cost of building has collapsed.

What is GStack, Garry Tan's toolkit?

GStack is an open-source toolkit Garry Tan built to turn Claude Code into an AI engineering team. It wraps a coding agent in roles, process, and review (what he calls a thin harness with fat skills) so the agent does structured work instead of one-shot guessing. It grew out of Tan automating his own repetitive workflow, including QA, and has more than 70,000 GitHub stars.

Does Garry Tan still write code as CEO of Y Combinator?

Yes. Tan built the first version of Bookface, YC's internal social platform and knowledge base, during his 2011 to 2015 stint as a partner, and after a multi-year hiatus he returned to hands-on building, shipping hundreds of thousands of lines of code with AI while running Y Combinator. He directs the tools himself rather than delegating everything, which is a large part of why founders study how he builds.

Garry Tan: The Founder Playbook

Q: Who is Garry Tan?

Garry Tan is the president and CEO of Y Combinator, a role he has held since January 2023 when he succeeded Geoff Ralston. Before running YC he co-founded the blogging platform Posterous (acquired by Twitter in 2012), was the tenth employee at Palantir where he led engineering on the finance product and designed the logo, and co-founded Initialized Capital, which he grew to about $3.2 billion in assets under management. He was an early investor in Coinbase and Instacart and wrote Coinbase's first seed check.

Garry Tan is the president and CEO of Y Combinator, the accelerator that has funded companies like Airbnb, Stripe, and Coinbase, and studying how Garry Tan builds is worth the time for any founder shipping with AI. He is not a career administrator who wandered into the job. He co-founded the blogging platform Posterous (a YC Summer 2008 company that Twitter acquired in 2012), was the tenth employee at Palantir where he led engineering on the finance product and designed the logo, then co-founded Initialized Capital and grew it to roughly $3.2 billion in assets under management before returning to run YC in January 2023. He wrote Coinbase's first seed check. What makes him worth copying, more than the resume, is that he still ships code, and he has turned how he does it into an open playbook that any of the tech founders in this roster can steal. You can follow his day-to-day thinking on his X account.

Run agents like a team, not a chat window

Tan's central move is to stop treating a coding agent as a genie you prompt once and start treating it as a team you manage. In How to Make Claude Code Your AI Engineering Team, Garry says "We are in a completely new era of building software, the agent era." The problem, in his telling, is that a raw agent has no structure. In the same walkthrough he explains why: "Out of the box, the model wanders. It doesn't know your data well, so it guesses."

His answer is GStack, an open-source toolkit (more than 70,000 GitHub stars) that wires Claude Code up with roles, process, and review, the same scaffolding a human engineering org runs on. He calls it a "thin harness, fat skills" approach: keep the orchestration light and load the real judgment into well-defined roles. His summary of the whole idea is that the way to get agents to do real work is the same way humans have always done it, as a team, with roles, with process, and with review.

So what: before you point an agent at your codebase, define the roles and the review step. An agent with a spec, a reviewer, and a definition of done produces work you can trust; an agent handed a one-line prompt produces a plausible guess you then have to debug.

Pressure-test the idea before you write a line of code

Tan built an "Office Hours" agent skill that copies what YC partners actually do in a session: it interrogates the product before any code exists. The centerpiece is a single forcing question. In How to Make Claude Code Your AI Engineering Team, Garry asks "What's the strongest evidence that you have that someone actually wants this?" The skill runs six of these forcing questions and reframes the product before a builder opens an editor.

This is the YC partner instinct encoded into a tool. The scarce resource in the agent era is no longer engineering hours, it is knowing what deserves to be built, so Tan spends the cheap step (an agent grilling him) to protect the expensive one (weeks of building the wrong thing).

So what: run office hours on yourself. Before you brief an agent to build a feature, force yourself to answer what evidence you have that anyone wants it. If the honest answer is thin, you just saved the build.

Stay a builder, even from the top job

Tan is the CEO of the most influential accelerator in tech, and he still writes software. He is proud of it. In How to Make Claude Code Your AI Engineering Team, Garry says "I actually built the first version of Bookface, YC's internal social platform and knowledge base," establishing that his instinct to build never went away. After a multi-year hiatus from hands-on coding, he came back and shipped hundreds of thousands of lines with AI while running YC.

The lesson for founders is that AI collapses the excuse that you are "too senior to build now." Tan directs the tools rather than delegating everything to a team and losing contact with the product. He stays close enough to the work to have opinions that are grounded in reality instead of status reports. Founders who model this stay dangerous.

So what: do not let your title pull you out of the code. Use agents to stay hands-on at a level you thought you had to give up. The founder who can still open the product and change it moves faster than the one who can only ask someone else to.

Automate the drudgery you personally feel

GStack did not start as a product. It started as Tan automating his own annoyance. In How to Make Claude Code Your AI Engineering Team, Garry says "I found myself sitting there doing QA. Probably the least fun part of software development." So he built QA and review into the toolkit, automated the part he hated, and then open-sourced the whole thing.

That is a repeatable pattern for finding what to automate: the best target is the repetitive work you personally do and personally dread. You feel the pain sharply enough to specify the fix precisely, which is exactly what an agent needs. Founders often try to automate glamorous work and miss the grind that is quietly eating their week. Tan did the opposite, and the byproduct (a widely used tool) followed from solving his own problem well.

So what: list the tasks you do every week that you dread. Point your agents there first. The drudgery you feel most acutely is the automation you can specify best, and often the thing other founders will want too.

Act on the timing, because the barrier just fell

Tan's framing for founders is that the constraint has moved. In How to Make Claude Code Your AI Engineering Team, Garry says "This is the most incredible time in history to build software," and he closes the walkthrough with "The barrier to building just collapsed. The only question left is what are you going to build?"

His point is not cheerleading. It is that the bottleneck used to be the cost of building, and now it is ambition and judgment about what to build. Tiny teams can now reach revenue per employee numbers that were impossible a few years ago, which is exactly the shift that lets a solo founder build like a team. The founders who win in this window are the ones who treat the collapse of build cost as a reason to aim higher, not as an excuse to ship more of the same.

So what: assume the build is cheap and spend your scarce judgment on the target. Pick the ambitious thing you would not have attempted when engineering was the bottleneck, then let the agents close the gap.

What to copy this week

Give your coding agent a role, a spec, and a review step before you point it at real work; never one-shot a vague prompt.
Run YC-style office hours on your own idea: demand the strongest evidence that someone actually wants it before any code gets written.
Get hands-on again yourself: use agents to make one change to your own product this week instead of only delegating it.
List the repetitive tasks you personally dread and automate the worst one first; that is where an agent pays off fastest.
Pick one target more ambitious than you would have attempted when engineering was the bottleneck, and start it now.

AI Operating System for Startups