AI-assisted QA agent

The problem

Most of a QA analyst’s time on a ticket is spent before any testing starts: finding the acceptance criteria, reading the code change, working out what else it might touch, checking what has been deployed and what regression already covers. That work is repetitive, and when it is rushed the test scope is too narrow.

What I built

An agent that does the gathering and the first pass of analysis:

• Context first: pulls the ticket, acceptance criteria and related technical context before looking at code.
• Change analysis: reviews the diff and estimates its blast radius across frontend, API and database layers.
• Status checks: looks at deployment state and existing regression coverage.
• Output: a structured, risk-based test plan including exploratory testing suggestions, which a human reviews and edits before use.

It lives in the engineering team’s shared AI tooling repo, so anyone on the team can run it against a ticket. It is built in Kiro from reusable steering files, hooks and specialist sub-agents, all running on Claude models. Smaller, benchmarked models handle the information-gathering steps to cut token usage and cost; a more capable model does the reasoning.

What I learned

• Routing tasks to the cheapest model that passes the benchmark changes the economics of running an agent in a real process.
• The agent is useful because a person still reviews its output; removing that step would have made it worse, not faster.