The idea that AI agents can develop software almost autonomously has been at the forefront of the tech debate for months. While some question whether the so-called software factories are truly viable, Cloudflare has presented a use case that’s already live in production. The company claims to have reduced the number of open issues in the Astro project from over 200 to just a few dozen thanks to an automated system based on AI agents, with the goal of achieving an incident-free repository for the first time in its history.
The key points of Cloudflare’s software factory in 20 seconds
- Cloudflare automates incident management for the Astro project using AI agents.
- The system reproduces errors, diagnoses them, generates fixes, and prepares test versions.
- The automation runs entirely via GitHub Actions.
- The open framework Flue enables building similar workflows for other projects.
- Astro has cut its open incidents from over 200 to about 30.
In recent months, Cloudflare has developed a chain of automation that goes far beyond simply responding to a ticket. Each incident follows a full flow that attempts to reproduce the issue, identify its cause, verify if it’s truly a bug versus intended behavior, and when possible, generate a ready-to-test solution for the reporter.
AI reproduces, analyzes, and fixes errors
The system starts when a user opens an issue on GitHub.
Instead of sending the query directly to a single language model, the platform divides the work among several specialized and independent agents.
Each performs a specific task:
- reproduce the problem;
- diagnose the root cause;
- check whether the behavior is actually an error or expected functionality;
- prepare a potential fix.
Each phase produces a report that serves as input for the next agent, preventing a single model from accumulating all the context or trying to solve the problem too early.
According to Cloudflare, this approach also mitigates one of the common issues with large language models: their tendency to assume there is always an error to fix.
GitHub Actions acts as the assembly line
All automation runs directly on GitHub Actions.
The workflow operates like a state machine governed by repository tags. Each incident automatically shifts status as the analysis and validation progress.
When the AI identifies a potential fix:
- it generates a preliminary version of the package;
- publishes the full analysis logs;
- asks the user to test the correction.
If the user confirms the problem is resolved, the system automatically creates an associated pull request.
From an internal tool to an open framework
During development, engineers realized that most of the system didn’t depend specifically on GitHub.
The logic was simply to react to an event, launch a sequence of independent agents, and coordinate their results.
This abstraction evolved into Flue, an open-source framework designed for building AI agent workflows that can run across multiple platforms, including GitHub, Slack, scheduled tasks, or any webhook-based system.
Its goal is to provide a reusable infrastructure for developing complex automations without being tied to a specific language model provider.
A new way of managing open source projects
One challenge Cloudflare aimed to address was the increasing workload faced by maintainers of open source projects.
The rise of AI tools has multiplied incidents, change requests, and even security reports, while the time to review them still depends on human effort.
The company states that automation has not only prevented burnout but also allowed contributors to focus on higher-value tasks such as technical discussions, reviewing new features, or collaborating directly with other developers.
When AI makes mistakes, it still improves the software
Cloudflare highlights that some of the system’s most interesting effects occur when agents err.
Usually, errors stem from issues like:
- insufficient documentation;
- unclear architecture;
- limited test coverage.
For instance, the company reports errors related to the Hot Module Replacement (HMR) system, where AI repeatedly proposed modifications that fixed specific bugs but introduced regressions elsewhere.
Adding explanatory comments and unit tests allowed the agents to stop repeating those mistakes, simultaneously enhancing overall understanding of the project for both AI and future developers.
A model that could extend to other projects
Cloudflare has made all this logic available as a standalone action called triagebot-action, open-source, and already adopted by other projects beyond Astro.
Though still evolving, the company sees it as a practical example of building software factories: chains of specialized agents capable of automating repetitive development tasks without eliminating human oversight.
The aim isn’t to replace developers but to reduce administrative overhead and allow engineers to focus on the core evolution of the software.
Frequently Asked Questions
What is an AI-based software factory?
An automated workflow where multiple AI agents perform different software development tasks, such as analyzing issues, generating code, or validating changes.
What is Flue?
Flue is the open-source framework developed by Cloudflare for building workflows based on autonomous AI agents that can operate across various platforms.
How does Cloudflare use AI in Astro?
AI reproduces errors, analyzes causes, generates potential fixes, publishes test versions, and automatically creates a pull request once the user confirms the fix works.
Has Cloudflare manually deleted old issues?
No. The company states that the backlog reduction was achieved not by closing inactive issues but by automating their analysis and resolution through AI agents.
