The talk link: https://x.com/poteto/status/2102050467505430555/video/1

  • Treat your codebase as your agent memory
  • Record decisions/rules the agent needs to follow inside the code package
  • Clean up tech debt — agents tend to copy the anti-patterns and short-term solutions in the codebase
  • Create verification skills to improve trust in agent work, so the agent can verify its own work
  • Identify the mistakes agents make and consolidate them into the codebase/static analysis (linter/compiler/CI)/rules (AGENTS.md)/skills/‘style guide’
  • How to verify UI work? How can we decouple the UI layer and the data layer?
  • Scale the agent in the loop — make each step an agent, and remove the human from the loop as trust in the agent increases