Study of ~3,000 repos ranks how people configure coding agents: start with AGENTS.md alone
JohnAlexander · x · 2026-09-13
An article on agent harness design, based on scanning nearly 3,000 real repositories, ranks the eight common ways to configure a coding agent. Key advice: start with just a context file written for a new teammate, name it AGENTS.md (all major tools read it; keep a two-line CLAUDE.md pointer), and only add the remaining configurations months later once the context file is solid. It also addresses classic failure modes like agents claiming tests passed and forgetting early instructions.
More from coding & agent
- YC Demo Day: everyone is just building domain-specific harnesses — amplifiedamp · 2026-09-13
- Hermes CLI user reports /redraw misbehaving over SSH in VSCode — AnuranBuilds · 2026-09-13
- GHOSTBOUND: open-source tool that lets AI agents see only allowlisted files from private repos — TheFreakinOG · 2026-09-13
- Hand-writing permissions for every AI agent is becoming the real production tax — Prestigious-Run-1954 · 2026-09-13
- BountyScout bot hourly-scans GitHub bounties and floods maintainers with AI applications — gleech · 2026-09-13
- Better AI 3D output costs more upfront but saves tokens and time vs. endless iteration — chaseleantj · 2026-09-13