SQLite draws a line: no agentic code, yes reproducible bugs

SQLite now has an AGENTS.md file in its repository. Simon Willison notes it is not mainly an internal development guide. It is a message for people pointing AI coding agents at the SQLite codebase. The policy says two things at once: SQLite does not accept pull requests without prior agreement and legal paperwork, and it does not accept agentic code.

SQLite told agents what it wants and what it does not: evidence of the problem, not an automatic patch

SQLite allows agent-assisted bug reports if they include a reproducible test case. A patch can document a possible direction, but SQLite maintainers may reimplement the change themselves. A recent commit removed the word "currently" from the sentence saying SQLite does not accept agentic code. That is not a minor edit. It makes the restriction permanent, not temporary.

Maintainer time is what agents cannot return

SQLite is not a hobby library. It is critical infrastructure running almost everywhere. For a project like that, code quality is only part of the problem. Legal origin, auditability, author reputation and maintainer time all matter.

An agent can confidently produce a patch. But the maintainer pays the verification cost: checking whether the bug is real, whether the patch breaks invariants, whether it is a hallucination and whether the project can even accept the contribution. SQLite's rule says: send evidence of the problem, not someone else's automatic code wrapped as a gift.

AGENTS.md becomes a public contract for automated contributors

AGENTS.md is becoming more than a prompt file. It can be a public contract for automated contributors: what is welcome, what is noise and what will be rejected before it burns maintainer time.

Open source projects will need this. As agents make bug filing and patch generation cheaper, low-quality submissions become easier too. Without clear rules, maintainers drown in well-formatted noise. SQLite was not the first: the project's forum was overwhelmed by AI-generated bug reports of uneven quality, which led them to set up a separate bug forum to manage the volume.

AGENTS.md as a template: agent as an analysis tool, not as an anonymous code author

Watch whether other infrastructure projects add similar policies. The important distinction will not be the file format, but the acceptance policy. Agent as a tool for reproduction and analysis, yes. Agent as an anonymous author of critical code, no.

Lilith's verdict

This is the grown-up answer to AI spam: do not ban everything, define what has value. Agent patch no, reproducible test yes. Maintainers protect time, quality and legal cleanliness at once.