FlowForge

Inspiration

Software delivery still has work after code is written: failed tests, review, release notes and incident triage. FlowForge explores how GitLab Duo Agent Platform can help, while keeping each action visible and supervised.

What it does

FlowForge includes four custom AI Catalog agents: Test Medic, Reviewer, Release Scribe and First Responder. Their prompts and tool selections are versioned in the public repository. A sample Python service and GitLab CI pipeline provide a small, reproducible test case.

The demonstrated workflow is supervised, not hands-off:

1. A seeded discount-calculation bug produced three failing tests.
2. Test Medic diagnosed the failures from the supplied job output and source, then created a branch, fix commit and merge request after approval of each tool action.
3. A real GitLab pipeline passed on the fix branch.
4. Reviewer reviewed the supplied diff and posted a review note with two non-blocking findings.
5. The fix was merged, a tag was created, and Release Scribe created a release-notes issue.

First Responder is configured for incident triage, but an end-to-end monitor run is not demonstrated. Automatic pipeline repair, event-triggered orchestration and automatic merging are intended extensions, not verified capabilities of this entry.

How we built it

We defined the agents in GitLab's AI Catalog, using scoped tool grants and versioned prompts. The repository contains the sample service, pytest tests, CI configuration and workflow design. The demo uses GitLab Duo chat and approval-gated branch, commit, MR, comment and issue tools.

In this chat context, agents could not read repository files or run tests directly. Source, diffs and job output were supplied to them. CI, rather than an agent's unsupported claim, verified the fix.

Autonomy level

Start Fresh Supervised. Each demonstrated agent tool action required an approval. Humans supplied context, approved actions, merged the fix and created the tag. No zero-click or hands-off loop is claimed.

Accomplishments

The evidence chain is a real failing pipeline, an agent-authored fix MR, a passing pipeline, a review note, merged code and a Scribe-written release-notes issue. The public repository and 99-second video show this supervised path.

Challenges and lessons

Tool availability in chat differed from the intended workflow. Keeping that limit explicit prevented us from presenting a supervised demo as an autonomous system. Test Medic also refused to make a fix it could not support until it had sufficient context.

What's next

Validate event-triggered execution, repository-context access and pipeline-repair flows before claiming unattended operation. Exercise First Responder against a seeded incident and add tests for the Reviewer's two non-blocking findings.

Built With

GitLab Duo Agent Platform, GitLab CI/CD, Python, Flask, pytest

Built With

• gitlab
• gitlab-duo-agent-platform
• python