I ran my daily workflow through an agentic IDE for 30 days. One-line result: more diffs, same review capacity — sometimes not enough.
Clear wins: boilerplate, test scaffolds, migration drafts, README drafts. Clear loss: domain review time grew. Agents ship plausible diffs that break invariants — especially around money and auth.
My rule is simple: agent output = draft. No merge without passing money, inventory, auth, and idempotency checks. Pasting the agent summary into the PR description isn't review.
I don't treat speed as the metric; reviewable, understandable diff count is. Weekly retro: how many agent-heavy PRs, any incidents — numbers, not vibes.