sqlite AGENTS.md

Origin: An "Instruction Manual" for AI

Five days ago, a file named AGENTS.md appeared in the repository of the renowned embedded database, SQLite. It's not internal development documentation, but a clear "code of conduct" aimed at developers who might use AI coding agents (like Devin, Cursor Composer, etc.) to analyze or modify SQLite's codebase. The core message is direct: SQLite does not accept unsolicited pull requests, especially those generated by AI agents. However, the project welcomes AI-generated bug reports that include a reproducible test case. A recent commit even removed the word "currently" to strengthen the stance of "not accepting." Simultaneously, the SQLite forums were being flooded with a barrage of AI-generated bug reports of varying quality, forcing the project to segregate them into a brand-new "SQLite Bug Forum."

Deconstruction: The Signals Behind the Rules

This is more than a technical statement; it's a clear signal. Maintainer D. Richard Hipp and the team are drawing a boundary with their actions. They distinguish between two types of AI output: code contributions and bug reports. For the former, they are cautious, even resistant, likely due to concerns about code quality, copyright (requiring contributions to be placed in the public domain), and the precious review time of maintainers. For the latter, they offer a conditional welcome—provided the report is high-quality and includes a reproducible test case, which effectively aids in problem diagnosis. Diverting bug reports to a separate forum is a pragmatic engineering management decision, aimed at protecting core developers from being overwhelmed by the information flood, ensuring that effective signals (real, high-quality bug reports) can be heard.

Trend Insight: From "Technical Adaptation" to "Social Contract"

This incident reveals a trend deeper than technology: the proliferation of AI coding agents is forcing open-source communities to shift from adaptation at the technical level to establishing new rules for community collaboration. In the past, discussions focused on "how to make AI better understand codebases" (RAG, fine-tuning, etc.). Now, the question becomes "how should human maintainers cope when AI can generate code and reports at scale?" SQLite's AGENTS.md is the prototype of this new "social contract." It explicitly tells users of AI agents: "You can use your tools, but you must play by our rules." This foreshadows that major open-source projects in the future may all need a similar "AI Collaboration Guide," clarifying the project's acceptable scope, format requirements, and quality thresholds for AI-generated content.

Practical Value: Insights for Developers and Project Maintainers

For developers using AI coding agents, this is a crucial reminder: Do not blindly submit AI output. Before submitting any contribution to an open-source project, you must read and follow its contribution guidelines; now, you might also need to specifically check for an AGENTS.md. You need to understand that the recipient of your AI agent's code or report is an exhausted human maintainer. Submitting low-quality or non-compliant work only adds to their burden and may even be closed outright. For open-source project maintainers, SQLite provides a reference template. If your project starts receiving a high volume of seemingly AI-generated, low-quality contributions or reports, consider following SQLite's lead by establishing clear rules and implementing分流管理 (分流管理: diversionary management). This could be a necessary measure to maintain community health and development efficiency.

Counter-intuitive: AI's "Help" Might Be a Burden

An angle that might be overlooked is that AI makes "contributing" incredibly easy, but this isn't necessarily good for the project. In the past, submitting a patch required a developer to deeply understand the code, write it manually, and test it—a process that itself acted as a quality filter. Now, AI can generate a seemingly plausible patch or bug report in minutes. This "low-friction" contribution, if lacking rigorous human review and contextual understanding, becomes pure noise and burden for maintainers. SQLite's case vividly illustrates that technological "enablement" without corresponding "governance" can backfire. AI amplifies human capabilities, but it also amplifies chaos. How to manage this exponentially growing stream of "contributions" brought by AI is a serious challenge that all successful open-source projects are about to face, or are already facing.