Garry Tan

Garry Tan is the president and chief executive of Y Combinator, co-founder of Initialized Capital, and an early investor in Coinbase and Instacart. A founder, engineer, and venture capitalist, he is developing open-source infrastructure that gives AI agents durable institutional memory and turns organizational processes into reusable software.

From startup founder to Y Combinator CEO

Tan studied computer engineering at Stanford, worked at Microsoft, and became an early engineer at Palantir, helping develop its financial-analysis product. He co-founded Posterous, a blogging platform in YC’s Summer 2008 batch that Twitter later acquired.

As a YC partner, Tan advised founders and built Bookface, the accelerator’s private founder network. He left in 2015 to focus on Initialized Capital, the venture firm he co-founded, where his early investments included Coinbase, Instacart, and Flexport. He returned to YC as president and chief executive in January 2023, resuming leadership of an institution he had previously experienced as a founder, investor, and software builder.

Building companies with institutional memory

Tan argues that companies using identical AI models can produce dramatically different results because their leverage depends on how they organize knowledge, assign responsibilities, and verify completed work. His AI Engineer World’s Fair keynote describes an organization whose finance, events, media, and engineering teams all encode operational knowledge for agents.

• GBrain and institutional memory. His open-source GBrain provides agents with persistent organizational context through retrieval, linked information, synthesized answers, citations, and permission-scoped access. Its purpose is to surface the specific history relevant to each task, including previous decisions and conversations.

• Agent skills as organizational infrastructure. Tan treats Markdown-based procedures as bounded roles, resolver tables as routing mechanisms, and targeted evaluations as performance checks. Successful workflows become reusable skills that the organization can refine instead of reconstructing instructions for every request.

• Context engineering and knowledge hygiene. Useful retrieval requires documenting and enriching institutional knowledge, linking related material, tracking provenance, resolving contradictions, and removing obsolete information. Without active curation, agents can retrieve stale or conflicting claims with unwarranted confidence.

• Deterministic computation alongside model judgment. Tan assigns interpretation and human preferences to language models while keeping calculations, constraints, and durable state in conventional software. Event seating illustrates the distinction: an agent can evaluate which attendees should meet while explicit code tracks assignments and enforces structural rules.

Tan has argued that existing systems of record must become AI harnesses or risk replacement. Through both YC’s startup requests and his own open-source work, he is pursuing the same opportunity: organizational software that remembers what a company knows and improves as its agents complete new work.