coding-agents

The same model family solved every hidden-test repair through Pi and Grok Build, but one route reported sixteen times the total tokens. Most of the gap was context, not answers.

The runtime got the reviews. The transferable part is how a 2,500-pull-request, agent-written repository keeps itself honest, and one of its rules caught a stale instruction in my own setup the same afternoon.

Deleting your agent's accumulated instructions is safest for the people who never tested them, and most dangerous for the people who did.

In April my benchmark said coding correctness is model-determined and the harness is a thin proxy. Two newer results show exactly where that claim breaks: the harness matters at the edge of what the model can do.

A 208 MB log, 59,356 retries, and zero LLM calls. A debugging story about what happens when the symptom lies about the cause.

The architect writes the tests. The implementer makes them pass. No prose specs, no circular validation.

I dispatched a 952-line monolithic CLI through my own coding-agent dispatch system to be refactored into seven modules. It worked. Notes on what self-bootstrap reveals about agent harness design.

I scraped 16,181 of Simon Willison's posts and analysed the 395 from 2026. An inflection in November 2025, GLM-5 closing the gap, and why the harness — not the model — is the competitive moat.

I benchmarked four AI coding harnesses on 12 tasks using the same model. The harness barely matters for correctness — it's all about the model.