QA Agent Platform: current generation over A2A

QA Agent Platform is the current generation of the QA agent, moved onto the company's internal agent platform and connected to it over A2A. A ticket entering QA is handed to the agent, which generates test cases, explores the feature in a real browser and has an LLM judge decide pass or fail with cited evidence. On a failure, the bug gets filed back into Jira. It has been proven end to end against a staging environment.

Runs are Temporal workflows behind FastAPI, with separate queues for browser work and model work and a read-write lock per test environment, so writers run alone and readers run together. Replay skips the model when nothing changed, and prompt caching and per-run budgets keep the rest of the cost in check. Review is confidence-based, with a human approving each batch. Test accounts are stored encrypted and leased per case, and secrets are redacted.

The code is hexagonal, with a layering test that enforces the boundaries, and all model traffic goes through a LiteLLM gateway. I integrated a teammate-built code-graph retrieval service for grounding, and the changes the platform needed are proposed upstream.