I've had some success asking the Antigravity agent to find flaky tests and propose fixes for them, then just running it in the background and reviewing what it finds. Upstreaming this to the repo so others can use it, since it includes helpful notes like `--runs_per_test` and leveraging random seeds as well as an iteration loop.

I opted not to have the agent do anything with PRs just yet, but if this is useful and we build confidence in it, we can explore that in the future.