There is a fair number of abandoned OSS with 1 or 2 major issues that prevent them from continuing to work and isn't too difficult to fix - just requires time. If users pooled context (issues, code, docs) and tokens, I'm inclined to believe an AI agent could resolve them.
Here's an example: hydroxide [0] is a ProtonMail to IMAP/SMTP bridge that no longer works due to upstream API changes [1]. Working implementations are available, just in a different language [2]. All an AI would have to do is reference the current implementation and the alternative implementation, then migrate over to the new API format. Perhaps use an mitmproxy MCP to verify that the requests made are identical. Overall, not a very difficult problem and I can probably do it myself given a few days.
Human engineering hours are quite expensive though and for such a low value task, I would assume an AI agent would be more cost effective.
Is anyone working on a similar problem/premise?
[0] https://github.com/emersion/hydroxide [1] https://github.com/emersion/hydroxide/issues/304 [2] https://github.com/opulentfox-29/protonmail-api-client/blob/master/src/protonmail/client.py
rvz•9mo ago
I don't think AI agents can generalize reliably to maintain complex open-source software, better than the authors. Those who say that are out to sell you an AI product or something.
The benchmark I would consider is the maintenance of the Linux kernel and AI is far from approaching to the level that it can maintain correct critical code that is trustworthy enough to being merged into mainline. But we'll see.
> Overall, not a very difficult problem and I can probably do it myself given a few days. Human engineering hours are quite expensive though and for such a low value task, I would assume an AI agent would be more cost effective.
Software maintenance is indeed expensive for a reason, however we still can't assume that an AI agent can help deal with autonomous maintenance on OSS projects; large or small or even in [0] and [2].
Technical debt is also a cost and many of these abandoned OSS projects don't even have any tests in the first place, including both 'hydroxide' and the protonmail client in [2].
So how do you know you if either project really works or not? Even if you use AI, what if it introduces a new bug in its implementation? How would you know beyond 'it works for me'?
In fact, autonomous AI agents just risk introducing more technical debt in OSS projects and increases that cost rather than reducing it.