Pete Koomen

Confidence 0.75 · 2 sources · last confirmed 2026-05-28

General Partner at Y Combinator and the architect of YC’s internal AI agent infrastructure. Previously co-founder of Optimizely (one of the first and most widely-adopted A/B testing platforms for apps and websites). Author of the [[Horseless Carriages|Horseless Carriages]] essay — a critique of AI-as-feature-inside-deterministic-software pattern (the Gmail email-writer as canonical example) which argues that AI-native software should look like agent wrapping deterministic tools rather than deterministic software wrapping AI.

At YC, Koomen led the year-long effort (started ~mid-2025) to build the firm’s internal AI agent infrastructure: a shared-tool-registry (≥350 tools as of May 2026), a shared-skill-registry with the skillify / DRY / MECE resolver pattern, broadcast-by-default agent conversations, a self-improving dream-cycle, and a single Postgres-substrate one database to rule them all approach (“BigTable for agents”). His framing of the multiplayer-harness gap (Lightcone podcast, 27 May 2026) is the wiki’s first explicit single-player → multiplayer harness-engineering progression — naming the team / org-scale problem that the existing single-player harness sources (Khan, Osmani, Chatterjee) don’t address.

Appears in this wiki via

• 2026-05-19-garg-yc-internal-ai-agent-evolves-itself — referenced as the architect of YC’s internal infrastructure (Garg’s 5-minute IC-vantage on the same project).
• 2026-05-27-koomen-yc-lightcone-inside-yc-ai-playbook — guest on the Lightcone podcast (~46:30); first-person architecture-and-organisation talk on YC’s internal AI infrastructure; names the multiplayer-harness gap, the $100K-token-budget leapfrog argument, the egalitarian + trust-by-default precondition, and the two-sentence-pitch skill as worked example of org-scale super-intelligence compounding.

Open questions

• The Horseless Carriages essay itself — is it worth ingesting as a separate kind: article source page? (It is the conceptual substrate for Koomen’s AI-as-building-layer framing.)
• Pete Koomen’s role at Optimizely before YC — is there a substantive thread connecting A/B-testing-platform design lessons to YC-internal-agent-infrastructure design? (Worth flagging as a follow-up if a second source addresses it.)