SQLite tells AI agents: bring the test, not the code

The central sentence is legal as much as technical. SQLite does not accept pull requests without prior agreement and/or accompanying paperwork that places the pull request in the public domain. That matches SQLite’s unusual position in software infrastructure: it is not just another repository where a patch is submitted and waits for review, but a tightly maintained component embedded across browsers, operating systems, apps and devices. AGENTS.md is a reminder that procedural gates do not disappear just because a model can now knock on them at scale.

The sharper boundary concerns code itself. The document says SQLite does not accept agentic code. Human SQLite developers may still review a concise, well-written pull request as a proof of concept before reimplementing the change themselves. In practice, that means a patch can serve as an explanation of an idea, but not as a finished contribution ready to merge.

This is not a blanket rejection of AI-assisted work. The file leaves room for agentic bug reports if they include a reproducible test case. Patches or pull requests that demonstrate a possible fix for documentation purposes are also welcomed. The distinction is narrow but important: the problem, test and sketch of a fix can be useful; responsibility for final code remains with the human maintainers. For a project like SQLite, that is not a cosmetic distinction.

The latest commit made the position even firmer. The word “currently” was removed from the sentence saying SQLite does not accept agentic code, with a commit message describing the change as strengthening the statement. That edit matters because it changes the policy’s posture. It no longer reads like a temporary pause while maintainers wait for tooling to improve. It reads like a standing line: the code intake path will not be reshaped around the pace of generative systems.

There is also a community-pressure story underneath it. Willison reports that the SQLite forum was being flooded with AI-generated material, which helps explain why the project chose a format that agents and their users are likely to find. The SQLite forum becomes a visible symptom of a broader shift: open-source maintainers are no longer filtering only human submissions of uneven quality, but also automated flows that can produce plausible text without understanding local norms, licensing, or maintenance cost.

AGENTS.md is therefore less a curiosity about SQLite and more a signal for agentic development. Projects will increasingly need to state what an agent may do, what can be submitted as supporting evidence, and where experimentation stops. SQLite’s answer is short and pointed: a bug with a reproducible test can enter the conversation; agent-written code does not enter the codebase.