Letting an agent self-improve its own harness daily: resumed runs 7→77, memory noise 94%→0
AccBalanced · x · 2026-11-11
theandrewsiah spent days hand-building a DeepSeek superagent to speed up his personal pi harness — then dropped it and just told an agent to make the harness faster, better, cheaper daily. It worked: a Coach's Top-5 dashboard shows resumed runs 7→77, memory noise 94%→0, goal-agent idle gaps 142→61 min, and test-grader agreement with human labels 80%→93% on 30 fresh cases. Pattern: turn repeated mistakes into code, gates, caps, confirms.
More from coding & agent
- SF developer meetup 'Who's Watching the Agents?' to cover agent observability — davidcrawshaw · 2026-11-11
- Supabase lands in Expo EAS CLI: one command wires a backend into every environment — dshukertjr · 2026-11-11
- Ex-Meta researcher predicts a hard ceiling of 150-200B tokens for the SWE market — DimitrisPapail · 2026-11-11
- ComfyUI Challenge Winner ComfyArt: Open-Source Fully Local AI Design Canvas — bfl_ai · 2026-11-11
- One Photo of a Toddler's Playroom Becomes a Racing Game, Then the Whole House — every · 2026-11-11
- Workshop moment: user tells coding agent 'you jsut fo it i odnt care' — bushibuilds · 2026-11-11