Teamday
← Back

Garry Tan: A Markdown File Is an Employee

a16z
agentsfuture-of-workenterprisestartupautomationbusinessinterview

Why Garry Tan Says a Markdown File Is an Employee

Garry Tan, President and CEO of Y Combinator, talks with a16z about founder psychology and what agentic coding does to companies. The Silicon Valley history is the warm-up. The useful part is his operating model: do a task once with an agent, then turn it into a reusable skill.

Skillify everything: Tan describes a loop for any business process. "A markdown file is an employee. And it's an employee that will do the job perfectly every single time." You do the task once, expensively, with many corrections. The agent's trace becomes a skill file plus code and tests, runs on a cron job, and later failures are bug fixes that stay fixed. He applies this to sales, marketing and support, and says his own gstack project was the same pattern for engineering.

Token-max, on purpose: Frontier labs still ration compute, so Tan says serious users run agents like Hermes Agent or OpenClaw and load 800,000 tokens per request. "It costs, I don't know, 50 or 100,000 dollars a year to use the agents at full strength." For a CEO or founder he thinks that is worth paying: "you get to live in 2028 today."

Agents as the middle layer: He cites Brex CEO Pedro Franceschi, whose agent reads meeting transcripts from direct reports two levels down. It surfaces what is broken and who is in conflict, so he can arrive at a meeting with three weeks of context and settle it. Tan's point is that companies fail when the business no longer fits in one person's head. Agents plus memory and retrieval remove that limit.

What breaks at scale: Large skill libraries need provenance and conflict management. When two facts disagree, the newer, better-sourced one wins, and a cron job sweeps for stale data. He also sees more 35-45 year old founders with experience winning, and cites companies going from zero to $15M ARR in about four months with two or three people and a few hundred skill files.

SaaS and the API line: "A pure per-seat SaaS thing, not totally clear it will exist in another 5 or 10 years." It works as a wedge, but only if it leads to a data or network-effect moat. He also describes a YC-backed company whose bug-report endpoint is built for agents and replies in real time with a triage decision and a workaround.

The white pill on speed: His counterintuitive claim is that humans are the bottleneck. "Society is way slower than you think. Government is way slower than you think. Every company in the world is way slower than you think." Microsoft cannot reorganize around loops, but a startup can, and every startup must.

5 Takeaways from Garry Tan on Running a Company With Agents

  • Turn every repeated task into a skill - Do it once well, capture the trace as markdown plus code plus tests, schedule it. Failures become permanent bug fixes.
  • Spend tokens like a founder, not a consumer - Full-strength agents with huge context are expensive but give the CEO a compounding advantage.
  • Put agents in the coordination layer - Reading transcripts, tracking dependencies and spotting conflicts is work middle management did lossily.
  • Invest in provenance and sweeps - Big skill and memory stores go wrong without source tracking and recurring accuracy checks.
  • Expect incumbents to be slow - The opening is real, but displacement will take years, not quarters.

What This Means for AI-Powered Organizations

Tan's model is close to how an agent team should work: a skill is the unit of employment, the trace is the training, and a schedule is the shift. The constraint he names is organizational, not model intelligence, which is why small teams that redesign around loops can outrun larger ones.