Make merging and recovery cheap

  • The guiding question is how extra agent work can reduce human attention: start agents earlier and let them proceed further before the next review. 29:32 29:53
  • Theo estimates only 10–15% of his tokens write code, with roughly 85–90% verifying it. He presents reducing the likelihood of a flawed PR from 5% to 1% as a useful target, not a measured result from this video. 30:20 30:48
  • Reduce risk both before and after merge: agents should verify changes before review, while deployment systems should detect failures and make reversions fast. His proposed recovery threshold is under 15 seconds. 30:48 31:17 32:56
  • A task may include implementation, tests, filing a PR, waiting for review, and fixing feedback across multiple rounds. His example takes over an hour, leaving time to start other independent tasks. 38:03 38:28

Delegate the problem before prescribing a fix

  • Bring the agent in when thinking starts. Theo describes spending three weeks considering a problem before an agent suggested a simple shortcut he had missed. 33:20 33:45
  • Supply the observed problem and ask for simpler approaches. Prescribing a premature solution can cause the agent to implement the wrong change and require repeated corrections. 34:05 34:29
  • For a reported UI bug, he now gives the screenshot and asks for a fix. If the agent fails, that signals a problem needing his investigation and direction. 34:55 35:20
  • His management analogy is to assign interesting problems and intervene when solutions repeatedly miss. The intended role for human reasoning is unresolved, difficult work and learning from successful solutions. 35:20 37:02 37:29
  • While one agent works, start the next task and return when results need attention. The goal is parallel execution with sequential human decisions. 37:51 38:28