The team behind Astro, a popular JavaScript framework, has slashed its open GitHub issues from more than 200 at the start of this year to around 20 at the time of writing, according to a report published by The New Stack on Tuesday. The team expects to hit zero open issues within the next month, which would be a first in the project's five-year history. Cloudflare, which acquired Astro's parent company in January, announced Tuesday that it's releasing the AI-powered triage system as open-source software for other maintainers to use.
The tool, called triagebot-action, operates as a GitHub Action and runs a four-stage pipeline when someone opens an issue: it reproduces the bug in a sandbox, diagnoses the root cause, verifies it's an actual bug, and attempts a fix, the report says. Each stage is managed by a separate AI agent, and the entire process is controlled by a state machine encoded in GitHub labels, so anyone can see exactly where a given issue sits and why. Adoption beyond Astro itself is still early—Cloudflare's workers-sdk team is prototyping it internally, but only Astro currently runs it in production, according to Fred Schott, Astro's co-founder and now a senior engineering manager at Cloudflare. Both the triage tool and the underlying agent framework behind it, called Flue, are built to run on any infrastructure with no dependency on Cloudflare-specific tooling.
Schott tells The New Stack that the "fix" step is the least essential part of the system—triaging an issue and handing the report to a human maintainer is "still a pretty great outcome" even without it. "Fix is always the hardest because the bar is so high," he says. "Where all of the other tasks are just 'can you complete your phase—reproduce the bug, diagnose the cause, etc'. 'Fix' is the one where it has to write code that is of high enough quality that it can be merged." According to Schott, triage came first because issue triage proved to be something agents got very good at "without falling into some of the traps of what they were bad at re: code quality," and because automating it without requiring a maintainer to manually trigger it would save the team a ton of time.
The report traces the problem back to a familiar pain point for open source maintainers: AI coding tools have made low-effort bug reports, unreviewed patches, and speculative vulnerability write-ups trivially easy to generate, but working out whether any of it holds up still takes a person just as long as it always did. The report explains that Schott's team watched where agents kept proving reliable and where they didn't, gradually realizing that issue triage was time-consuming for maintainers and a task agents could handle well. Cloudflare's recent open-source releases suggest a broader shift: triagebot-action is the second tool in a week that Cloudflare has released with AI agents as the primary user, following pvcli, a command-line debugger that was "explicitly designed for agent-based debugging."
Looking ahead, Schott says he doesn't know when the system will reach the point where its fixes can be trusted close to 100% of the time, but that it's built to evolve toward that gradually. It launched internally with no fix capability at all, then added a suggested fix left on a branch for a reviewer to look at, and only later gained the ability to open a pull request directly from that suggestion. He can picture it going further still: the agent judging its own confidence or the size of a fix, and eventually opening or even merging pull requests on its own. "It's great when it works, but it's also always at the discretion of a maintainer for if it is correct and worth accepting, or better to go and write the fix yourself," Schott says.