Astro Automated Issue Triage With AI Agents, Flue Framework

An open-source team built a working software factory

The Astro project has reduced its open issue count from over 200 to approximately 30 through an automated triage pipeline powered by AI agents. The system reproduces reported bugs in sandboxes, diagnoses root causes, and generates preview releases for verification—all running inside GitHub Actions.

The achievement stands out in a landscape where maintainer burnout has intensified as AI-generated issues flood repositories. Rather than declaring issue bankruptcy or auto-closing stale tickets, the Astro team automated the time-consuming work of issue triage itself. The repository is on track to reach zero open issues for the first time in its five-year history.

How the triage pipeline works

The automation mirrors manual issue resolution through four distinct phases, each handled by an isolated AI subagent. First, the system clones provided reproduction repositories to verify reported issues. Next, it instruments the codebase with logging to pinpoint root causes. A verification phase reviews test suites and documentation to determine whether behavior represents a genuine bug or intended functionality. Finally, the system converts reproductions into failing unit tests and identifies appropriate fixes.

Each subagent operates independently and passes findings forward by compiling discoveries into a report file. This isolation prevents the common LLM tendency to force solutions when no bug actually exists.

The entire pipeline functions as a state machine driven by issue labels. New submissions receive a "triage needed" label, and the system reads through existing comments to determine current status and next actions. When agents identify a fix, the pipeline creates a preview release using pkg.pr.new and posts summaries, full logs, and installation instructions back to the issue. Once reporters confirm fixes work, the automation opens linked pull requests.

From project-specific tool to open framework

The underlying engine evolved into Flue, an open framework for building agent automation that works identically regardless of deployment context—whether triggered by GitHub issues, Slack messages, cron jobs, or webhooks. The team also released triagebot-action, a standalone GitHub Action that other projects can deploy directly or fork as a reference implementation.

The Astro team initially worried automated responses might feel impersonal. Instead, maintainers report more meaningful interactions with their community, now focused on Discord discussions, RFC reviews, and contributor collaboration rather than backlog administration.

Why it matters

This represents one of the first documented cases of AI agent automation delivering sustained results on a production open-source project. The approach treats agent failures as signals pointing to architectural problems, missing documentation, or insufficient test coverage—issues that affect human developers equally. When the triage bot repeatedly attempted incorrect fixes for Hot Module Replacement bugs, adding a descriptive comment explaining the logic stopped the errors and improved the codebase for future human contributors.

The system demonstrates that agent automation can reduce maintainer burden without sacrificing code quality or community engagement. As AI-generated contributions continue increasing repository noise, this model offers a practical path forward for resource-constrained open-source teams.

These details were first reported by the Astro team on the Cloudflare blog.