What it does:
A developer transitions a Jira issue → Jira Automation fires → a three-phase Rovo Agent pipeline analyzes the issue, assesses risk, plans test coverage, and writes full Gherkin scenarios → a custom Node.js service normalizes the output, converts it to Confluence storage format, auto-organizes it into the correct project/year/version hierarchy (auto creates them if non-existing), publishes the TAR, and links it back to the source Jira issue.
Minutes later, a fully structured TAR is sitting in Confluence with expand macros, data tables, risk scores, and a remote link back to the ticket. It's been running in production across 6+ projects.
The hard part nobody talks about:
Getting AI to generate test cases is easy. Getting it to think like a QA analyst is hard. The breakthrough was a three-phase pipeline where each phase is scope-locked so it refuses to do another phase's work:
Phase 1 — Requirements Analysis. Reads the Jira issue, decomposes it, identifies boundaries, parameters, and dependencies. Outputs structured analysis only, no test cases.
Phase 2 — Coverage Planning. Takes Phase 1's output, runs a 17-question risk assessment, scores the feature, and builds a Coverage Matrix. Here's the key, the risk score controls how many scenarios Phase 3 writes. Higher risk = deeper coverage. Lower risk = lean coverage. This is how real QA teams think, and getting the AI to dynamically calibrate its own output depth based on self-generated risk was the trickiest part of the whole build.
Phase 3 — Scenario Writing. Writes Gherkin scenarios against the Coverage Matrix, then reconciles to make sure nothing was dropped. No scenario exists without a Coverage Matrix entry. No matrix entry exists without a scenario.
Why the scope locks matter:
Without them, the AI drifts. It starts writing test cases during analysis, skips planning, and drops coverage. The scope locks force the plan-then-write discipline that makes the output actually reliable.
The recursive part:
I'm not a full-stack developer by trade. I used AI to help me build the Node.js publishing service, the markdown-to-Confluence conversion, the hierarchy routing, the version conflict handling. So the workflow is AI-built infrastructure running AI agents orchestrated by Jira Automation. The whole thing is a layer cake of AI leverage.
What I'd love to see from Atlassian:
Native Rovo Agent chaining, right now I'm orchestrating multi-phase pipelines through Jira Automation, which works but feels like a workaround for what should be a first-class feature
Richer Rovo to Confluence publishing capabilities so the custom service layer becomes less necessary
Agent-to-agent communication within Rovo itself
Happy to answer questions about the architecture, the prompt engineering, or the publishing service. This will genuinely change how our QA team operates; we went from hours of manual test analysis per feature to minutes of automated generation with human review.
