A week with Devin: honest notes

We handed Devin a week of our real backlog. Nothing staged, nothing cherry-picked — same Linear tickets a teammate would pick up on Monday.

What worked

• Dependency bumps and straightforward refactors. Devin shipped four of these cleanly. Read the ticket, made the change, ran the tests, opened the PR. Human review was 10 minutes each.
• Fixing clear bugs with reproducible steps. Reproduce, fix, test. Three tickets, all merged.

What didn't

• Tickets that were actually design decisions in disguise. Devin picked a direction and committed to it. Reviewer spent more time unwinding the picked direction than it would have taken to write from scratch.
• Long-running tasks that needed judgment mid-way. No mechanism to pause and ask. Either it blazes through a wrong assumption or stalls.
• Anything touching unfamiliar subsystems. Devin's context was narrower than we hoped. It kept re-reading the same files it had already seen.

The honest verdict

Devin is real. It is not a teammate yet. It's a pattern-specific force-multiplier that needs the same kind of framing good junior engineers get: "here's the clear shape of the work." Give it anything more ambiguous and you're the one paying the iteration cost.