• Poteto traces his agent workflow to a side project after leaving Meta's React team. He found himself micromanaging one agent, iterating on skills without a good way to measure their impact, and began extracting his own coding workflows for agents to follow. 2:12 3:03 3:50
  • After joining Cursor in March, he worked on performance problems in the Agents window. At first he manually used flame graphs and heap snapshots, then recognized that he was acting as the proxy between the agent and Chrome DevTools. 4:22 4:52 5:18
  • His skills aim to make the correct, rigorous approach the easy approach because even strong models tend to take shortcuts. 5:46 6:07
  • Poteto argues that domain expertise matters more as agents improve. The bottleneck shifts toward clearly expressing intent and goals; a domain expert who can use agents can build useful products without being a software specialist. 6:51 7:44
  • Both speakers describe language as a way to steer agent reasoning. Precise terms such as TDD or "tautological tests" compress a richer process or critique into language the agent can reuse. 8:11 9:08 9:34