I was recently quoted in The New Stack’s article on coding agents ignoring open source contribution guidelines, which covers research from Peking University highlighting a growing challenge for open source maintainers.
TL;DR: Coding agents are focused on completing tasks, not necessarily following a project’s contribution rules. Better instructions can help, but maintainers shouldn’t have to duplicate documentation across formats or rely on agents to police themselves. Existing review and contribution workflows remain essential.
The researchers tested four AI models against contribution policies across 49 repositories and found that agents rarely sought out those policies on their own. Even when explicitly reminded, they continued to contribute to projects that prohibited AI-generated code.
One point I raised in the article is that adding an AGENTS.md files isn’t necessarily the solution because too often these files duplicate existing contributor documentation, creating another place maintainers need to keep docs updated.
I’ve been experimenting with maintaining a single local AGENTS.md file and carrying those instructions across the coding tools I use (e.g. into Claude’s CLAUDE.md file). It’s not perfect, but it helps establish consistent expectations about discovering and respecting repository-specific guidelines.
My takeaway is that we need to focus less on creating additional documentation and more on making existing policies discoverable, consistent, and enforceable.
Leave a Reply