Steve Yegge says coding-agent harnesses should be bespoke, not reusable
dbreunig · x · 2026-08-04
Coding-agent harnesses should be bespoke, Steve Yegge argues
In this long post, Yegge says the key to making coding agents work on massive problems overnight is to think in terms of loops and graphs, not one-shot prompts.
- He argues that reusable harness frameworks are a dead end; harnesses need to be chemically bonded into the application.
- His own harness, Wheelhouse, is closed-source and purpose-built for his workflow.
- He says Gas Town was supposed to be reusable, but ended up being used to build itself, and then fell apart with Opus 4.7 because the model kept wanting to “just two more things” instead of converging on actual work.
- Since Claude Fable 5 arrived, he says he has gone back full-time to his 30-year-old game Wyvern, which the new model is finally smart enough to help him build.
The core thesis: agent harnesses are becoming bespoke infrastructure, and the right one can turn a project into something closer to a living city than a script runner.
Related event: Steve Yegge on AI Coding Agent Architecture Evolution(2 posts)→
More from coding & agent
- ThursdAI explores whether Codex and multi-agent math can solve Erdős problems — thursdai_pod · 2026-08-04
- Digibee used Opik to version prompts and catch agent regressions before users did — dl_weekly · 2026-08-04
- Stripe built its internal Kai knowledge agent in one week with one engineer and Deep Agents — LangChain · 2026-08-04
- Using Claude with prebuilt shaders and assets to generate game worlds — mathemagic1an · 2026-08-04
- How I AI episode shows a workflow for handing voice, web and video tasks to Codex — nickbaumann_ · 2026-08-04
- Benchmark paper says some agent evals spend over $3,000 per task to elicit more capability — dejavucoder · 2026-08-04