AI Agents Automate Astro Bug Triage, Cutting Open Issues by 85%

Cloudflare has deployed a system of specialized AI agents to handle issue triage for the Astro open-source web framework, dramatically shrinking its bug backlog. The automation achieved an approximate 85% reduction in open issues, dropping from more than 200 to about 30—with the team now aiming to reach zero.

How the Workflow Works#

Rather than running all agents in a single context, Cloudflare segmented the process into distinct stages, each performing its own specialized task. A reproduction agent validates that reported bugs actually occur as described, a diagnosis agent instruments the code to pinpoint root causes, a verification agent cross-checks findings against existing tests and documentation, and a fix agent converts problem cases into test coverage before proposing a solution. Agents communicate by passing structured information through a report file rather than sharing a unified execution state.

GitHub issue labels drive the entire system as a state machine. New reports start tagged triage needed, and as agents make progress, labels shift toward fix verified once the reporter confirms a proposed patch works. When a potential fix is identified, the workflow automatically builds a preview release, posts diagnostic logs and installation steps to the issue, and waits for human validation. Upon confirmation, it opens a pull request to integrate the fix.

Learning from Failures#

Cloudflare treats agent failures as design signals about the codebase itself. In one instance involving Hot Module Replacement, an agent repeatedly modified a conditional in ways that introduced regressions because the underlying behavior lacked adequate test coverage. Adding a descriptive code comment changed how the agent approached the problem and stopped the cycle of faulty modifications—demonstrating that improving test coverage and code clarity can make automation work better.

Beyond Astro: Broader Frameworks#

The Astro triage system became the foundation for two broader contributions: triagebot-action, a reusable GitHub Action, and Flue, an open-source framework for building long-lived agent workflows. Rather than requiring developers to write orchestration loops manually, Flue uses a declarative model where engineers define an agent's context (model, skills, sandbox, and instructions) upfront. Workflows persist their execution history via an append-only event log, enabling interrupted work to resume from where it left off.