Descriptions:
Y Combinator general partner Pete Kumman — founder of Optimizely and the architect of YC’s internal AI infrastructure — joins YC’s Gary Tan for the first public discussion of how one of the world’s most influential startup accelerators has rebuilt itself as an AI-native organization. Kumman describes how a project that started roughly a year ago with automating finance workflows has since grown into a full agent infrastructure layer used across YC’s entire operation.
A central concept in the episode is YC’s ‘skills’ abstraction — a layer on top of tools that lets non-engineers encode their own workflows and share institutional knowledge across the organization. Kumman explains how the system has evolved from manually written system prompts to meta-prompting, where agents write and improve their own skills. He details a nightly ‘dream cycle’ in which a general agent reviews all employee-agent conversation transcripts, identifies gaps, and autonomously updates context and skills — drawing parallels to similar loops at Andrej Karpathy’s work and OpenAI’s Codex.
The episode also covers how YC applies this infrastructure to partner workflows, including the ‘two sentence description’ coaching skill used with founders, and frames the broader lesson as context engineering: helping people replicate their internal knowledge in others’ minds. Essential viewing for anyone building or scaling AI agent systems inside an organization.
📺 Source: Y Combinator · Published May 27, 2026
🏷️ Format: Deep Dive
