Manus vs Cursor vs Devin: which AI agent actually ships production code?

Three coding agents, three weeks, the same six tasks. Manus, Cursor, Devin. Here is what actually happened.

The tasks 1. Refactor a 600-line component into smaller modules without changing behaviour 2. Add a webhook receiver to an existing Express app with HMAC verification 3. Migrate a Postgres table without downtime (rename column, backfill) 4. Write integration tests for a Stripe checkout flow 5. Debug a memory leak in a Node.js service running in Docker 6. Build a small CLI tool from scratch (200 lines, with tests)

Cursor Wins on tasks 1 and 4. Cursor's file-tree-aware suggestions catch context that the others miss. As a pair-programming tool, it is unmatched.

Loses on tasks 3 and 5. Cursor wants to apply changes immediately, not investigate. For debugging, it suggests rather than diagnoses.

Devin Wins on task 6. Building from scratch with a clear spec is what Devin is best at.

Loses on tasks 1, 2 and 5. Devin can take 40 minutes on tasks the others finish in 10. Cost per task is also significantly higher.

Manus Wins on tasks 2, 3, and 5. Where the work is "investigate first, then act", Manus is consistently the best of the three. The investigation phase is where it pulls ahead: it does not start changing things until it has read the relevant code.

Loses on task 4 (Cursor was much faster) and matches Devin on task 6.

My takeaway Cursor: stays open in my editor at all times. Best for in-flow work. Devin: useful for "spec a thing, walk away, come back". Pricey. Manus: the one I reach for when I need a non-trivial task done end-to-end without supervision. The free credits on signup are enough to test it. [Sign up here](https://manus.im/invitation/AIRTDVWVEWKCK4R) to try it yourself.

Caveats Three weeks is not a long sample. Your codebase shape matters: agents that work on a clean greenfield Next.js app fall apart on a 200,000-line legacy Java codebase.