AI issue triage cut Astro's backlog 85% — and it's open source

Most "AI cleared our backlog" claims are a screenshot and a vibe. This one has a number you can check. On August 4, Cloudflare published how it drove Astro's GitHub issue count toward zero using AI issue triage agents — from over 200 open issues down to roughly 30, an 85% cut, with the maintainers expecting zero inside a month. That would be the first time the repo has hit zero in its five-plus year history.

What actually happened

The interesting part isn't the model. It's the pipeline shape.

Incoming bug reports run through four stages: reproduce the bug in a sandbox, diagnose the root cause with instrumentation, verify the behavior is actually a bug, and fix it by converting the issue into a failing test.

Each stage is executed by a separate, isolated subagent. They don't share a conversation. They hand off by writing findings into a report.md file that the next stage reads. That design exists for one reason: a single long-running LLM session will talk itself into fixing a bug that doesn't exist, because it started with the premise that there is one. Isolation gives the verify step a chance to say no.

The models are Cloudflare Workers AI — @cf/moonshotai/kimi-k2.7-code for triage, kimi-k2.6 for verification. Preview releases go out through pkg.pr.new so the person who filed the issue can confirm the fix on their own project before anything merges. The engine is open-sourced as Flue, plus a standalone triagebot-action GitHub Action.

Humans didn't leave. Cloudflare's framing is that maintainers stopped administering a backlog and moved to RFC discussions and new feature requests.

Why AI issue triage matters for your business

You probably don't maintain an open-source framework. You do have a queue that never empties — support tickets, warranty claims, intake forms, bug reports from your own staff. The pattern transfers directly.

Three things to copy. Split the job into stages with separate agents. One prompt that reads, judges, and resolves will resolve everything, because agreeing is cheaper than disagreeing. Make the handoff a file, not a chat history. A written artifact between steps is auditable; a context window isn't. Give the reporter a verifiable output. A preview build the customer can test is a different kind of proof than a ticket marked resolved.

The honest caveat: a code repo is an unusually good fit. Bugs reproduce deterministically, tests pass or fail, and the ground truth is machine-checkable. Your refund queue has no test suite. Before you automate a queue, ask what the pass/fail signal is. If you can't name it, you're building a machine that closes tickets, not one that resolves them.

Key takeaways

Astro's open issues went from 200+ to ~30 (85%) using staged triage agents; the framework is open source as Flue, with a triagebot-action GitHub Action
Four isolated subagents — reproduce, diagnose, verify, fix — hand off through a report.md instead of a shared context window
Isolation is the guardrail: it stops the model from inventing a fix for a bug that was never real
Runs on Workers AI (kimi-k2.7-code / kimi-k2.6), with preview builds via pkg.pr.new so the reporter verifies before merge
Copy the pattern only where you have a real pass/fail signal — otherwise you've automated ticket-closing, not resolution

Every business has one queue that quietly eats a day a week. We build staged agent pipelines with verification steps you can audit, not black boxes that mark things done — see how we build, or tell us which queue is bleeding.

Sources: Cloudflare Blog, The New Stack.

AI issue triage cut Astro's backlog 85% — and it's open source

What actually happened

Why AI issue triage matters for your business

Keep reading

Valar Atomics raises $1B: your AI bill is a power bill

Your AI footprint is 3x what your model inventory shows