Issue Scout

Inspiration

Developers (also Development agents) waste time ramping up on every new issue — figuring out which files matter, what patterns to follow, and what's been tried before. We wanted that first investigation step to happen automatically the moment an issue is assigned.

What it does

When an issue is assigned, Issue Scout explores the codebase, reads the issue and its comments, searches for prior art in related issues and MRs and posts a structured comment with entry points, patterns, risks, and a suggested approach.

How we built it

Pure YAML on the GitLab Duo Agent Platform — a single AgentComponent flow with 11 tools (search, file reading, issue/MR lookup, and comment posting). No code, no dependencies, just prompt engineering and flow configuration.

Challenges we ran into

Getting the agent to target the correct project was our biggest hurdle. Numeric project IDs return 404 on the platform — we had to hard code the project URL in the system prompt. We also started with a two-agent design (scout + reporter) that added complexity without value and collapsed it into one.

Accomplishments that we're proud of

It works end-to-end with zero code — just YAML and prompts. The output is genuinely useful: specific file paths, real patterns from the codebase and concrete steps rather than generic advice.

What we learned

Simpler flows are better flows. Our two-component design introduced context-passing bugs that a single agent avoided entirely. We also learned that the platform's tool resolution is strict about project identifiers — a lesson we learned the hard way.

What's next for Issue Scout

Adding support for multi-repo projects, smarter prioritization of which files to read (token budget awareness), and triggering on MR creation to scout review context for reviewers.

Gitlab Feedback

• It's very difficult to tell the Flow what the Project ID is.
• When a Flow runs into an error, there is no way to check logs for debugging.

Built With

• claude