Garry's List

Pete Koomen and Garry Tan explain how YC built internal agent infrastructure around a shared database, a tool registry, and agent harnesses, enabling teams to use AI for real work. They describe how transcripts and shared agent conversations can capture employees’ expertise and improve reusable skills, turning organizational knowledge into a shared resource. The talk argues that AI works best when employees can shape the prompts and tools themselves, and that open, high-trust organizations can use it to help people learn and work more effectively. It also makes the case for small, adaptable software and personal AI that people can customize, rather than tightly controlled systems.

Garry Tan describes using AI coding agents to build and maintain software at a much higher pace, including rebuilding a blogging platform in days and creating tools such as G Stack and G Brain. He explains tokenmaxxing: spending more on models and supplying broad context to make research and coding more complete. His workflow combines detailed Markdown skills, planning and review, automated tests, browser based QA, and a human who supplies goals and judgment. He argues that builders should learn to repair brittle agent setups and keep control of their own tools and prompts as personal AI becomes more capable.