What they did
Namduri, a QA engineer, built a learning project called QualityHub AI, an app that links requirements, test cases, test runs and defects. She began one task at a time with Cursor and ChatGPT, then moved to Claude Code. Once back at a full-time job she wrote down her QA philosophy and a set of change levels that decide how much review each change needs, pre-approved everyday work, and kept anything risky, costly, irreversible or undecided for herself. She then set up seven agents (Manager, Product Owner, Scrum Master, Architect, Developer, QA and Documentation) to run the project, reaching her through Claude notifications on her phone. When the team said it was ready, she ordered an exploratory test session with a charter written by the QA agent, and later added an outside reviewer agent.
How it went
Phone pings started at a couple a day and fell as the written rules answered more questions. Walking the demo herself she found a dead end the scripted tests never hit: a test run created without test cases could not get any added. The exploratory session found 17 bugs that all automated tests had missed, mostly half-working saves and silent data loss; no agent could reach another user's data. The team fixed 16 the same day, each with a regression test, and she shipped the last as a known issue. The outside reviewer caught overstated claims in her article draft. Her attempt to teach the QA agent her judgment did not work well.
Worth knowing
Green tests and an agent saying 'ready' were not enough here; the real bugs came from deliberately leaving the happy path. The value came from writing the rules down first and piloting them on small tasks. She does not describe the exact multi-agent setup or what it cost.