Are AI coding agents actually saving time, or just moving it to code review?

With Claude Code, Codex and Cursor’s agent mode all getting better at multi-file changes, a lot of teams now let agents open full pull requests.

What I keep hearing is that writing code got faster, but review got slower. Agent PRs tend to be bigger, touch more files and need more careful reading, because the reviewer didn’t write any of it.

So what does this mean for an existing stack?

• Do you limit agents to small, well-scoped tasks?
• Have you changed your review process (smaller PRs, required tests, a second reviewer)?
• Are you measuring time from task to merge, or just time to first commit?

Would love to hear from teams that have used agents for a few months. Did total cycle time actually go down?