Garry Tan on Taste, Agents and Founder Ambition

Do What You Know Not What’s Hot
- Do work you uniquely know and care about instead of chasing what’s hot.
- Garry Tan regrets leaving web programming in 2003 and urges founders to follow direct experience, not headlines or investors’ maps. (Time 0:02:35)
The Palantir Dinner He Said No To
- Garry turned down a chance to join Palantir despite Peter Thiel offering a $70k check and flying him down for dinner.
- He calls it a multi-billion dollar mistake and a lesson to trust people you know and the territory below the map. (Time 0:07:30)
How YC Turned Outsiders Into Insiders
- YC standardized access to Silicon Valley by using a simple application and letting builders be judged on ideas and execution.
- Garry credits Paul Graham and Jessica Livingston for opening doors to founders who lacked pedigree or networks. (Time 0:11:12)
Agents Multiply Founder Leverage
- The startup game changed: one person with agents can become 400x more productive, so founders should be far more ambitious.
- Garry warns pure per-seat SaaS without data or network moats may not survive the next 5–10 years. (Time 0:14:35)
Learn AI By Building Trivial Things
- Try trivial projects and use every model to build a chassis for learning the new AI technology.
- Garry says ship small consumer experiments and record office hours to build product intuition quickly. (Time 0:16:48)
Markdown Files Become Employees
- Garry frames reusable agent skills as codified operational knowledge: a Markdown file is an employee.
- He converts repetitive processes into skill files, tests, and cron jobs that an agent executes perfectly after iterations. (Time 0:21:11)
Pay To TokenMax For High‑Value Agents
- Token‑max on critical agent workflows when it makes sense: pay for large context and memory to operate at higher ‘IQ’.
- Garry notes top‑tier agent setups can cost tens of thousands per year but let founders ‘live in 2028 today.’ (Time 0:22:21)
Recursive Improvement With Skill Files
- Agentic systems create recursive improvement loops: automate a task, fix bugs, convert fixes into permanent skills, then scale to many agents.
- Garry claims two people plus hundreds of agents can reach $15M ARR in months using hundreds of skill files. (Time 0:24:31)
Provenance Is Critical As Agents Scale
- Scaling agentic operations raises provenance and conflict management needs; provenance decides which fact wins.
- Garry says cron jobs and trace history become crucial to resolve divergent agent outputs and keep accuracy. (Time 0:26:18)
Agents as Midlevel Managers
- Mid‑level management tasks can become agentic: agents can read meeting transcripts, detect conflicts, and surface root causes.
- Pedro at Brex built monitoring agents that flag team issues and let him walk into meetings already briefed. (Time 0:28:16)
The Next Computer Will Be A Memory‑First Voice Assistant
- The next computer will emphasize voice and persistent memory, becoming a benevolent assistant knowing your hopes and context.
- Garry predicts diarization, long memory, and voice will shift how people interact long term, not immediately. (Time 0:41:21)
From Jury Duty To Local Civic Action
- Garry got engaged in local politics after seeing failures in San Francisco institutions during COVID and serving jury duty.
- He launched GarysList.org to restore accountability on issues like elder crime and algebra access in schools. (Time 0:44:46)