We have had recent papers, Skillsbench, and then this one that indicate automatically generated Agents.md actually reduced performance. For projects creating automatically generating files, like onboarding, it would be nice to show we are building on these research, using good design principles, and can validate some improvements over not having automatically generated files. - https://arxiv.org/pdf/2602.11988v1

Yeah, I agree! One of the things that we’re hoping to look at is if we can use past sessions to improve agents.md, like we did for skills in our blog post. https://openhands.dev/blog/20260227-creating-effective-agent-skills

It'll be interesting to see what signals matter for improving AGENTS.md files. That paper indicates decreased resolution rate on SWE-bench style tasks (although I'm still a little skeptical of those results), but also indicates that it improves agent usage of repo-specific tools like

uv

. Personal pet peeve of mine is when OpenHands decides to muck up my system Python install running tests when it should just use

uv

as indicated in docs and workflows.