What you get: a written review of your tools, prompts and data flow, ranked by risk and by cost. A plan for safe writes: the agent proposes, a separate step applies, and every change is logged and reversible. Cost reductions, usually fewer and narrower tools and smaller tool results. The first fix implemented as a pull request, with tests that run against mocks so no live credentials are needed.