AI UAT Support Agent
A human-in-the-loop agent that investigates UAT defects the way a senior engineer would.
The problem
Every UAT defect meant a developer rebuilding context by hand — the Jira thread, the QA screen recording, logs scattered across services, what was actually deployed when it broke — before writing a single line of fix. Slow, and expensive in AI tokens.
What it does
- Pulls the Jira defect, comments and attachments, and reads the QA screen recording frame by frame.
- Reads the real code and runs git blame / log before suspecting a line: is this old, stable code, or did something just change?
- Queries Grafana Loki across services and follows a single request end-to-end.
- Checks Helm history to see which commit of each service was live when the incident happened.
- Builds a case, not a guess: every theory may only cite evidence it actually gathered.
- ✓00:01jiraPulled VO-49712 — 3 comments, 1 screen recording
- ✓00:04videoExtracted 14 frames → error toast at 00:41
- ✓00:09codeTraced submit handler → proposal-service
- ✓00:12gitblame: line unchanged for 14 months → look for what changed
- ✓00:15helmproposal-service bumped 2 days before the report
- ✓00:23lokiFollowed trace id across 4 services → 422 from rules-engine
- ✓00:31theoryPayload field renamed in new release; cites 3 pieces of evidence
- ›00:34gateProposed 2 approaches — awaiting developer approval ▍
Illustrative replay — not real ticket data.
Key decisions
- One gate
- It never edits code until a developer approves an approach.
- Zero infrastructure
- Runs inside the developer's own agent session. No backend, no hosted state.
- Evidence-only reasoning
- Logs, traces and video frames are tagged; conclusions must cite them.
- Clean hand-offs
- Investigations live in a shared Git repo, so anyone can
git pulland resume.
- 3× faster UAT defect resolution
- 2× lower AI cost per ticket
- <30 min smoke-test fix
