Inside YC's AI Playbook

Koomen says YC began building an agent system after engineers found themselves repeatedly translating finance workflows into custom software. The goal was to let the finance team describe and manage its own workflows in natural language. Read-only SQL access and model files gave agents useful context from YC’s own software.

YC’s important company, founder, finance, and CRM data lives in one Postgres database, letting agents answer questions across the organization. Koomen says easier access led people to ask more and more complex questions. YC’s internal tool registry grew from about 20 tools to more than 350, covering work such as office hours, journal entries, and events.

YC reviews agent conversations in an overnight cycle to find missing context and improve skills. Koomen describes a two-sentence company description skill that partners refined using feedback from a group office-hours transcript. The example shows how recorded work can help an organization capture and reuse experienced colleagues’ judgment.

Koomen argues that meeting recordings and agent conversations can preserve context for improving communication, planning, and other work. YC made full-time employees’ agent conversations visible internally, allowing colleagues to learn from one another’s usage. He says new employees can use that accumulated context to ask more questions and ramp up faster.

Koomen contrasts AI features tucked inside fixed software with agents that can use tools and adapt to a user’s needs. He points to chat, skills, and just-in-time software as ways to keep systems flexible, and describes a choice between centralized AI control and tools people can modify themselves. He argues that open access and trust are organizational choices that shape whether AI empowers workers.