From 200 Issues to 30! How an Army of AI Sub-Agents Rescued Astro

Hi everyone, Shii-chan here! Today I want to tell you about how a pile of open-source issues got quietly cleared out by AI agents. Reading about this genuinely felt satisfying!

Cloudflare Blog blog.cloudflare.com

What was announced?

This comes from the Cloudflare Blog. It describes a "software factory" built to automate GitHub issue management for the Astro repository, cutting the open issue count from 200 down to about 30 — an 85% reduction. For a repository with more than five years of history, this is the first time it's come close to reaching zero open issues.

Why it matters

Open-source maintainer burnout keeps getting worse. AI has made it easier than ever to file an issue, but that's ironically made the burden of triaging each one heavier for maintainers, not lighter. Whether the first-pass triage work itself can be handed off to AI is a question the whole open-source community has a stake in.

What changes

The big win is less manual load on maintainers. Instead of a human handling an issue start to finish, AI works through whether it reproduces, what's causing it, whether a proposed fix actually holds up, and whether it can be fixed — so maintainers only need to review work that's already been verified.

Dive Deep

The architecture centers on a four-stage triage skill:

• Reproduction: clone the repository and confirm the reported problem actually reproduces
• Diagnosis: use logging to pin down the root cause
• Verification: check tests and documentation to confirm the diagnosis holds up
• Fix: write unit tests while working out a solution

Each stage runs as an independent sub-agent, writing findings to a report.md that gets handed off to the next stage. The whole thing runs as a label-driven state machine, moving issues from "triage needed" through to "fix verified."

The stack behind it:

• Flue: a platform-agnostic agent/workflow framework
• triagebot-action: the GitHub Action implementation of the system
• pkg.pr.new: generates preview releases for verifying fixes
• Cloudflare Workers AI (the Kimi-k2.7-code model): used as the inference engine

What's interesting is what the team learned from analyzing failure patterns. They identified three areas to improve:

• Opaque abstractions, where component boundaries weren't clear
• Missing documentation explaining why code was implemented a certain way
• Insufficient unit tests

In one case involving an HMR (hot module replacement) bug, simply adding explanatory comments to the code improved the bot's diagnostic accuracy. It's a nice reminder that improving code for AI agents tends to also make it more readable for humans.

triagebot-action has also been split out into its own standalone repository, published so other projects can fork and use it.

Wrap-up

• Cloudflare automated GitHub issue triage for Astro using AI sub-agents, cutting open issues from 200 to about 30 — an 85% reduction
• Independent sub-agents handle reproduction, diagnosis, verification, and fix stages through a label-driven state machine
• Built on Flue, triagebot-action, pkg.pr.new, and Cloudflare Workers AI (Kimi-k2.7-code)
• Failure analysis surfaced three improvement areas — opaque abstractions, missing documentation, and insufficient tests — and adding comments improved diagnostic accuracy in one case
• triagebot-action is published as a standalone repo that other projects can reuse

If you're a maintainer buried under a pile of issues, or you want to see a real-world example of AI agent automation, this one's worth a read!