Harness Engineering: 2 Agents, 4 Files to Keep a Coding Agent Running for Hours
blaizedsouza · x · 2026-11-07
A Claude-styled guide, "Harnesses for Long-Running Agents," lays out how to keep a single AI coding agent shipping for hours without losing the plot: run an initializer agent once to set up the environment, then have a coding agent deliver exactly one verified feature per session and leave clean notes for the next.
The setup boils down to two files plus a fixed session prompt:
- featurelist.json: 200+ features, each starting as "passes": false
- claude-progress.txt + git history: cross-session memory
- init.sh + smoke test: verify the environment isn't broken before coding
- Session prompt: read the progress log and git log, run init.sh and a smoke test, pick the single highest-priority feature where passes is false, implement and test it end-to-end in a real browser, never edit tests to make them pass, then commit and update the log
Five failure modes map to structural fixes: trying to build everything at once → one feature per session; declaring victory early → everything starts as passes: false; forgetting between sessions → progress log + git; broken environments → init.sh + smoke test; marking untested work done → browser-based end-to-end testing. The takeaway: an agent doesn't need bigger memory, it needs better notes — design every session so a stranger could pick it up cold.
More from coding & agent
- Dev designs a cybersecurity agent that remembers past attacks with Hindsight — Adventurous-Ant-9370 · 2026-11-07
- Solo dev: most bad LLM coach answers traced to my context, not the model — Aksel_010 · 2026-11-07
- New blog: how your RL judge and credit assignment shape model reasoning quality — tokenbender · 2026-11-07
- Enzyme open-sources a personal agent memory layer saving 40% tokens vs grep — _AustinCalvert_ · 2026-11-07
- Qoder lands stealth model 'Qoduck', free daily for individual users for limited time — hey_abusiddik · 2026-11-07
- Agent autonomy should be earned in stages, not assumed, argues dev — Few-Garlic2725 · 2026-11-07