Reported throughput

  • Pocock says he has four agents working on separate projects with long-term goals, while he checks their plans and questions their approach. He is also experimenting with a chief-of-staff skill for sub-agent management. 31:04 31:25
  • He estimates roughly 200–300 commits a day, compared with about 20 on a previously good day. These are self-reported activity counts, not measurements of useful output or code quality. 31:25 31:45

His model and effort preference at recording time

  • Pocock names Claude Code and a model the transcript calls "Opus 5.5." These notes preserve the transcript's model label. 31:45
  • He reports using medium effort and considering lower effort for faster responses and better token efficiency. He attributes his ability to do this to an environment where tests catch mistakes; this is his workflow preference, not a general effort-level benchmark. 32:08 32:29
  • By harness, he means the application surrounding a model, giving Claude Code, Claude.ai, and Codex as examples. He thinks improvements to these tools matter alongside improvements to the model. 29:44

The concrete improvement to his video editor

  • Previously, the agent could not reliably run and inspect the application. Local OBS and other third-party dependencies complicated access while Pocock was also recording with it. 33:02 33:29
  • After fixing the setup, the agent could click through the application, capture screenshots, and run it properly. Pocock reports a large improvement in output quality because the agent had better evidence about the application. No numerical quality gain is given. 33:29
  • He prioritizes improving this working environment over continually comparing models. His suggestion that an older model could perform similarly with the improved setup is a hypothesis, not a demonstrated comparison. 33:29 34:03