Cloudflare has automated issue triage for the Astro open source framework using isolated AI agents running in GitHub Actions. The workflow reproduces reported bugs, diagnoses root causes, verifies behavior, and proposes fixes before generating preview releases for reporters to validate. Cloudflare reports that Astro's open issue count fell from more than 200 to about 30, an approximately 85% reduction based on those figures, with the team targeting zero open issues.
The workflow mirrors the steps Astro maintainers followed during manual issue resolution. Each stage runs as a separate subagent. A reproduction agent verifies the reported behavior, a diagnosis agent instruments the code to identify the cause, a verification agent examines tests, documentation, and comments, and a fix agent converts the reproduction into tests before implementing a solution. Agents pass information through a report.md file rather than sharing a single execution context.
Astro AI issue triage workflow (Source: Cloudflare Blog Post)
Cloudflare implemented the workflow as a state machine driven by GitHub issue labels. New issues receive a triage needed label, while confirmation of a proposed fix moves an issue toward fix verified. When an agent identifies a potential fix, the workflow creates a preview release and posts findings, logs, and installation instructions to the issue. After the reporter validates the patch, the automation opens a pull request. A July 2026 Astro issue involving the Container API, for example, was marked triage: fix verified after the reporter confirmed the bot's fix.
Thrives wrote in a LinkedIn comment that the significance is not simply having AI verify an issue, but running agent work in a sandbox so human reviewers primarily see results that have passed through the automated process. Jordan Matthiesen, Senior Product Manager at CloudBees, similarly highlighted reproduction before diagnosis and making fixes easy for reporters to test, while Shubhanshu Singh described the workflow as an example of explicit agent system design.
Jordan Matthiesen, Senior Product Manager at CloudBees, highlighted the importance of reproducing an issue before attempting a diagnosis, making fixes easy for reporters to test, and improving code, tests, and documentation when agents encounter difficulties. Shubhanshu Singh similarly described the Astro workflow as an example of structuring agentic systems through explicit system design rather than relying primarily on agent loop abstractions.
Astro AI issue triage workflow (Source: Cloudflare Blog Post)
Cloudflare also treats failed agent runs as signals about codebase maintainability. In one Hot Module Replacement case, an agent repeatedly modified a conditional and introduced regressions because the behavior lacked sufficient tests. Adding a descriptive code comment changed the agent's behavior and prevented the repeated modification.
The Astro workflow later became triagebot-action, a standalone GitHub Action, while its orchestration model evolved into Flue, an open-source framework for building durable agent workflows. Flue uses a declarative model in which developers define an agent's context, including its model, skills, sandbox, and instructions, rather than writing an orchestration loop. Its execution history is persisted through an append-only event log, allowing an interrupted workflow to resume from its previous state.
Flue can integrate agents with GitHub, Slack, Linear, and Discord and can run on Node.js, GitHub Actions, or Cloudflare infrastructure. On Cloudflare, agents can run as Durable Objects with durable execution and isolated storage. The Astro triage workflow therefore provides one example of a broader Flue model in which bounded agent tasks, persistent state, external events, and human approval points form a durable software workflow.