Keep pulling the thread on Pete Kooman.
A key capability for Y Combinator's internal AI agent was a tool that allowed it to run read-only SQL queries against the company's production database.
Y Combinator's ability to effectively use AI agents is significantly enhanced because all of its important operational data resides in a single Postgres database.
Pete Kooman believes the AI industry has not yet solved the challenge of a "multiplayer harness" that would enable agentic capabilities at a team or organizational level.
Y Combinator's internal AI infrastructure includes a shared tool registry that has grown from an initial 20 tools to over 350 tools today.
Y Combinator has a general agent that runs nightly to review all employee agent conversations and identify potential improvements and missing context.
Gary Tan asserts that building "superintelligence" within a company requires using artifacts like meeting transcripts to continuously and automatically improve operational skills via meta-prompting.
Pete Kooman argues that the true potential of AI is to shift control of software from the developer to the user, a concept he explored in his essay "Horseless Carriages."
Pete Kooman predicts that future AI-native software will primarily consist of an agent wrapping deterministic tools, rather than deterministic software wrapping an AI model.
Gary Tan believes a critical choice must be made in the next 18 to 24 months between a centralized AI future controlled by a few large companies and a decentralized, personal AI future.
Gary Tan advocates for a personal AI revolution where individuals can control their own agents, prompts, and models, rather than being subject to the platforms of Meta, Alphabet, OpenAI, or Anthropic.
Pete Kooman created all of the agent infrastructure at Y Combinator, including the harnesses for internal AI use.
Since the launch of ChatGPT, Y Combinator has primarily focused its funding on AI companies.