Instead of playing whack-a-mole with junk code, these maintainers are building what they call "software factories." They aren't waiting for a human to submit a fix; they're deploying their own LLM agents to handle the entire lifecycle of an issue.
The Vercel "Software Factory" approach #
Vercel is probably the best case study here. Their AI SDK project was hitting a breaking point with over 1,000 open issues and nearly 800 pull requests sitting in the backlog by mid-year. Instead of hiring a small army of engineers to triage, they deployed a multi-agent system.
They don't just use one "smart" bot. They use specialized agents for specific tasks:
Bug Reproduction Agent: Takes an issue report and tries to recreate the failure in a sandbox.Fix Implementation Agent: Writes the actual code to resolve the bug.Review Agent: Checks the proposed fix against the codebase to ensure it doesn't break anything else.
The logic is pretty cold but practical: they trust their own highly-optimized prompts more than they trust a random agent running on a contributor's laptop. As Vercel engineer Lars Grammel put it, if you've tuned an agent specifically for a certain category of bugs, you have a level of confidence in its output that you just can't get from community-driven PRs.
The results are actually insane. After just four weeks of running this deployment architecture (which includes custom UIs, APIs, and sandboxed execution environments), Vercel reported that their "factory" was authoring between 25% and 35% of all merged PRs and closing 70-80% of incoming issues.
How Astro flipped the script on backlogs #
Astro is seeing a similar shift. Fred Schott, the creator, mentioned that for five years, issues were coming in faster than any human team could possibly manage. They were stuck in a permanent state of "trimming the backlog" rather than actually building features.
By implementing an auto-triage system, they've moved to a model where agents handle:
-
Initial triage and labeling.
-
Reproduction of the reported bug.
-
Prompting the user to verify a suggested fix before a human ever looks at it.
This has fundamentally changed their workflow. Instead of a mountain of debt, they treat issues as a weekly priority list. It's a complete shift from reactive maintenance to proactive, agent-led management. It’s actually what led Schott to build Flue, a framework designed specifically around this idea of controlled, agentic contribution.
If you're managing a large-scale open source project, the move toward a private "software factory" might be the only way to survive the flood of AI-generated code hitting GitHub every single day.
[Stop bloating open source repos with AI-generated junk just to 6d ago](/en/news/7888/)
[LLM-generated spam is basically a DDoS attack on open-source 8d ago](/en/news/7685/)
**A $13B price tag for the AI developer playground everyone 9d ago
Developers are losing their minds over a new open-source model 10d ago Why AI can render a perfect Mario but fails at a simple vacuum 10d ago
MCP server runs on Android 14d ago Next Google is rolling out the Fairwind Program to give governments →