Harness, not model, moves coding agent from 30th to top 5 on Terminal-Bench
blaizedsouza · x · 2026-09-13
Paul Iusztin (Decoding AI) launched an open-source course, Building a Coding Agent From Scratch, arguing the harness — not the model — determines coding agent quality. In LangChain's Terminal-Bench experiment, changing only the harness moved an agent from 30th place into the top 5 with the same model.
Key points:
- The core agent loop is tiny: Reason → Act → Observe → Repeat. Real engineering lives around it.
- Production concerns: context delivery and compression, tool access and approvals, sandbox and subagent execution, memory persistence, tracing/debugging/evals, result delivery.
- The course builds Decode lesson by lesson: bare loop → sandboxing → context engineering → subagents → remote headless durability → evals and replays → a cloud swarm of parallel agents.
- Designed after studying how Claude Code (via leaked source) and OpenCode actually work, helping readers understand Claude Code and Codex under the hood.
More from coding & agent
- Dev ditches code editors: terminal plus Codex or Grok CLI is all you need — Kuprel · 2026-09-13
- Dev burns 1.6B tokens building a full playable game with GPT Astra in 4 days — bennash · 2026-09-13
- Rime's Coda voice model generates natural conversational AI voices in under 60 seconds — dr_cintas · 2026-09-13
- Human-in-the-loop isn't human authority: scoped grants beat click-approval fatigue — arthaudm · 2026-09-13
- New survey maps recursive self-improvement in agents by autonomy stages — dair_ai · 2026-09-13
- Orgtree v2: multi-agent orchestrator rebuilt as self-contained Electron app with parallel accounts — DynaBeast · 2026-09-13