BLOG
Research, techniques, and field notes on building for agents.
AI Agents Don't Fail. They Stop Too Early.
I spent two weeks building context tools for AI coding agents. The data killed every hypothesis — then revealed the actual problem. Here's the full story of how sourcebook check was born.
Fast Agents Are Just Agents That Don't Wander
We ran 19 bug-fix tasks with repeat runs. The biggest finding wasn't speed — it was variance. Same task, same bug: handwritten context produced 21s to 252s. sourcebook produced 34s to 57s. A map, not a boost.
We Scanned 30+ Repos. Here's What Broke (And What We Fixed).
esbuild, vLLM, Polars, Biome, and 25 more. 12 scanner bugs found, 69 QA assertions added, and an engine that stopped hallucinating about codebases.
What We Found Scanning 15 Open-Source Repos With 100,000+ Files
Next.js, Cal.com, Supabase, Django, and 10 more — hub files, invisible coupling, and the traps agents fall into.
We Benchmarked AI Context Files on Real GitHub Issues. Handwritten Briefs Won. Then We Caught Up.
We tested four context strategies on real codebases — no context, handwritten briefs, repo dumps, and sourcebook. Here's what happened when we measured real patches on real issues.
Why auto-generated CLAUDE.md files make your AI agents worse
ETH Zurich research shows auto-generated context files hurt performance by 2-3%. The only context that helps is what agents can't figure out alone. Three techniques that extract what they miss.