Garry Tan on YC, Founder Regret and the Coming 'Harness Wars'

Garry Tan opens with a story from his Microsoft years: he and his PM mentor had to walk over to the Windows team because that team would not answer emails or fix bugs. The bat was a prop, not a plan — his mentor brought it, and Tan describes it as a vague threat rather than anything anyone intended to follow through on.

He turns the anecdote into a binary. An organization like Microsoft cannot solve a problem that way. A startup can. And, in his framing, every startup must.

Tan graduated from Stanford in 2003, just as web 1.0 ended and the NASDAQ crashed. He had two offers: Microsoft's Windows Mobile group, which genuinely excited him, and Expedia. He took Microsoft because he had student loans to pay.

What he did not see was that leaving web programming was a mistake built into the timing. Mark Zuckerberg was doing FaceMash around that same year; Facebook did not exist yet. Tan's counterfactual is concrete — five years of web experience stacked on top of what everyone else had would have been an end run into building almost any kind of social software.

He was 22 and describes himself as being whipped around by whatever he could read in the Wall Street Journal. The question he now hears from founders — what's hot, what should I work on — he calls the wrong question entirely.

The right question, in his telling, is what are you interested in and what do you know uniquely. That is the frame he applies both to his own 2003 decision and to the founders asking him for a market read.

His own summary of that period is that he was getting out of web programming at exactly the wrong time because it felt like the smart thing to do.

Acharya puts a name to the undercurrent in the conversation: being outside the light cone, a permanent underclass — a phrase Tan himself had used earlier. Tan does not dispute the feeling. There is simply more competition than ever, he says, and then adds the part he thinks is actually actionable: people are spending too much time in distribution.

He borrows Chris Dixon's idea maze framing to explain where the good material is. The cool parts are the weird fringe parts, and Dixon's measure of interestingness is whether you can find the edge at all — for the most interesting things, Tan says, you typically cannot even find it.

The pattern he keeps returning to is that what looks like a toy is what comes to pass, and the people who build it are not the ones everybody is watching. His example is the home computer era: the Homebrew Computer Club was not where you were getting laid, and the people who built that world, in his description, were fringe weirdo punks and outsiders.

Asked about Palantir — lower status a decade ago, either because people were nervous about national-interest work or dismissed it as a services business — Tan says he has a worse story. Joe Lonsdale and Stefan Cohen were his fraternity brothers at Stanford and interns at Peter Thiel's hedge fund.

Thiel's hiring practice, as Tan describes it: write down all the smartest people you need to hire and go get them. They flew Tan down from Seattle for dinner.

Thiel offered him $70,000 — exactly what Tan was making at Microsoft at the time. Tan thanked him and declined, explaining that he might get promoted to level 60 that year. He did get promoted.

The framing he puts on it now is a $2 billion to $4 billion mistake. In the retelling, that promotion is the punchline, not the achievement.

Before YC, Tan says, you had to go to the right schools and know the right people and try to get invited to the right parties. He describes a social network that had to be cracked, and notes YC has a pejorative for that whole scene: sceners.

The mechanics were deliberately unglamorous. The original submission form was a website with 12 questions; the one-minute video came later. Today YC has 16 equal partners, including Tan, all of whom went through YC and had successful outcomes. The pitch to applicants is that builders on the other end will try to give you a fair shake.

At a TechCrunch event, Tan asks, are you really going to be able to tell anyone anything real? Founders hold their cards close, and he mocks the scripted exchange: how you doing, killing it, I'm also killing it.

Founder life, as he describes it, is solitary at specific moments rather than in general — the day you lose your best customer, the day your best engineer quits.

He also names the case he considers worst: the day your co-founder loses hope and you are not allowed to.

Acharya's claim is that code is no longer precious — the gauntlet he and Tan had to run just to ship something is gone. That, he argues, is why you used to need a sense that what you were making was important with a capital I.

The problem is that this requirement flies against earnestness, because earnestness is not obviously important with a capital I, and most important things start out looking unimportant.

Tan agrees with a short 'oh yeah.' The exchange sets up the question Acharya presses later: if the barriers are gone, why is the same small slice of people still doing the work?

That becomes the question for Tan. Can agency be embraced and practiced, or are some people naturally high agency and most people not?

Tan's answer is personal. He calls himself a late bloomer.

The evidence he offers is his own list of mistakes: declining Palantir, and leaving web programming to go learn. His conclusion is that his life is somewhat a testament to making mistakes, introspecting on them, reasoning about them, and making different choices afterward.

Tan's starting observation is that the frontier labs let you use ChatGPT or Claude but still constrain you on cost and on how much compute they hand over. To token max, he says, you need something like a Hermes agent or openclaw and to tune it all the way up — loading a million or 800,000 tokens per request, kept in something like soul.md.

Doing that is what lets you, in his phrase, live in 2028. He puts the cost at roughly 50 or $100,000 a year to run agents at full strength at 150 requests. The return, he argues, is felt almost immediately, and for a CEO or founder the math makes sense — the hard part is giving yourself permission to token max in the first place.

Then comes the loop. After a feat of strength, you turn it into a markdown file plus code plus tests that can be reused and dropped into a cron job. The application is any business process: do it once perfectly, badly and expensively the first time, and every future screw-up is just a bug fix. He names the domains — sales, marketing, customer support, everything.

Does the system need error correction? Do the errors compound? Does it get constrained by exception handling and human relationships? His premise is that otherwise these companies should be able to grow and run forever.

Tan's answer is that once you have a lot of markdown files, or a large amount of data, provenance and conflict management start to matter.

That is a narrower failure mode than the one Acharya proposed. It is not human relationships that bind the loop in Tan's telling — it is knowing which file is right when two of them disagree.

Tan's claim is that every business in the world is available for you to take once you have this capability in your brain. The reason is a limit on the other side of the table: the world is built around human beings who can only keep seven plus or minus two things in their head at any given time.

You plus an agent, in his formulation, can keep roughly three Harry Potter books in your head. That asymmetry is the opening.

He notes how many crypto-isms applied ten times more to AI, including the line that this is the worst it will ever be. All business, all the things people rely on, government especially, is still built on a world predicated on limited human beings, and it is not ready for a world where agents can think at the speed of light with unlimited amounts of data.

What he wants from that opening is not cheaper products. Products should be ten or a hundred times better — he says a thousand times better — and services too. It is not about lowering the cost of something, it is about creating magical things.

Tan returns to Redmond with detail. He and his PM mentor had to walk across the highway to the Windows team because that team would not reply to emails or fix bugs — they would not even mark the bug as won't fix. They simply ignored it.

He estimates they spent about four hours trying to get a P3 bug fixed that was a blocker for a thousand users, with seven levels between his team and the person who could actually fix it. Acharya frames the episode as a waste of humanity and 2 a.m. anxious wake-ups caused by someone avoiding a hard conversation.

Tan recalls doing competitive research for Windows Mobile versus the iPod after he got an iPod. Today he imagines the same job as a markdown file and a skill on a cron that maps who is blocked, by whom, and what the external dependencies are. A release that took six months, he says, could happen in a day or two.

His claim is about who can multiply it out. A big organization like Microsoft cannot; a startup can — and every startup must be organized this way. He proposes that the middle layer of bureaucracy be agents: the executive team sets direction, individuals decide execution, and the organizational layer in between should not be human.

Two references back this up for him. One is Venkatesh Rao's idea of being above the API line — an Uber driver is below it, and should fight to get above it — and Tan says AI erases that line. The other is Claw Visor, a YC-funded API company whose bug-report endpoint is built for agents rather than humans and replies in real time with something like: that's a good bug, we'll fix it next week, try this workaround.

The person doing the work on the line is empowered to change the line, which is what Tan cited as the reason Japanese cars got good and American cars started to struggle.

Acharya attributes roughly a 40-year generational run to that process power — the payoff came from process, not from a single product.

The comparison sits at the hinge of the argument. Tan's proposal that agents fill the middle layer of a company is a claim about process, and the precedent he reaches for is a manufacturing system built on letting the worker change the line.

Tan's expectation is that the cool thing is actually going to take longer than people now assume, and that nobody is even talking about this yet. His guess is that next year, 2027, will be the harness wars.

The blocker he names is cost: it is just too expensive right now. He predicts that today's frontier model compute will cost maybe 50 or 100 bucks in two or three years, and that the price drop brings back a war for a billion consumers — the browser wars will be back on. Acharya agrees repeatedly through the prediction.

Acharya sets up the closing turn: the earlier part of the conversation was utopian and about intelligence-bound problems, and this part is about problems bound up in human coordination.

Tan's white pill has two parts. The first is believing that America still has rule of law. The second is that people in government are trying to help and are somewhat earnest — not perfect, as he puts it.

His concrete example is jury duty in San Francisco. What he saw was an orderly process drawn from a real cross-section of the city, and it produced a realization about his own industry's place in it: we are a big part of the city, we're 10, 20% of the people, directly working in tech.

He also singles out the judge, calling the judge really good, though it was a horrible case. His conclusion is that government is not here to ruin people.