Most AI tools I work with write code: the feature, the fix, even the tests. Far fewer help with the question that comes first: what in this change actually needs testing, and how?

Reviewers look at code quality. QA gets a card that might have the information they need: description, acceptance criteria, definition of done, and tons of notes in the discussion thread.

This session shows how to build a small set of reusable skills and agents that examine a PR and return a test strategy, not code. The output names the critical paths the change touches. It also says whether each path calls for manual, smoke, or automated testing, and why.

We'll walk through how the work is split up: one agent analyzes the diff, one assesses risk, and one recommends the test strategy. We'll also cover the classification rubric that drives the recommendations and the guardrails that keep agents advising instead of authoring. You'll leave with a pattern you can use to reexamine the Software Development Lifecycle.