When Cognition launched Devin in March 2024, the internet collectively lost its mind. “The first AI software engineer.” A demo showing it building and deploying full applications autonomously, while mainstream media ran “AI will replace programmers” headlines for weeks straight.
Then people actually used it, and the vibes shifted hard. Now that Devin has been in the wild for almost a year, we pulled five real tasks off our own backlog and made it earn its keep. Spoiler: it is neither the end of programming nor a scam — it is something far more mundane, and far more interesting.
What Devin Actually Is
Devin is an autonomous AI coding agent that operates in a sandboxed environment with a browser, a code editor, a terminal, and a planner. You give it a task in natural language; it breaks the task into steps, writes code, runs it, debugs errors, and iterates until it declares victory.
How it differs from Copilot and Claude Code
Copilot and Claude Code are assistive tools — they work alongside you, and you steer constantly. Devin is designed to work independently. You assign a task, go make a coffee, and come back to a pull request.
That distinction is the entire product. It is also, as you are about to see, the entire problem.
The pitch vs the product
The launch demo showed Devin shipping full applications end to end with no human in sight. The pitch was a colleague. The product, as we found, is closer to a very literal-minded contractor who never asks a clarifying question — and that gap is where this whole review lives.
What It's Actually Like to Use
The five tasks we gave it
No synthetic benchmarks, no to-do apps. Five real tickets from our backlog: add a date range filter to an existing API endpoint, build a React component matching a Figma design, write a data migration script for a Supabase database, fix a bug in a webhook handler, and set up a new Astro page using existing component patterns.
Task one: the API filter — genuinely good
Devin read the existing codebase, understood the pattern, and added the filter with proper validation and tests. The code was clean and the PR was mergeable with minor comments. About what you would expect from a competent junior developer, which is quietly impressive for a robot.
Task two: the Figma component — rough
Devin can write JSX, but matching a specific design requires visual understanding it simply does not have. The structure was right; the spacing, typography, and responsive behaviour were all wrong. A human would have spent less time building it from scratch than fixing this output.
Task three: the migration script — partial credit
The script worked. It also missed edge cases in the data that any developer familiar with the project would have caught immediately. That is the scariest failure mode in this whole review: code that runs green right up until it quietly mangles production data.
Task four: the webhook bug — hard fail
Devin spent 45 minutes going in circles, trying fixes that never addressed the root cause. The bug required understanding the interaction between two services, and Devin could not hold that mental model in its head. It just kept iterating, confidently, on the wrong theory.
Task five: the Astro page — the best result
Following existing patterns is exactly what Devin excels at, and the output here was production-ready. Point it at a codebase with strong conventions and it colours inside the lines beautifully.
Where Devin Shines
The pattern across all five tasks is clear. Devin is excellent at well-defined work with existing patterns to follow: add a CRUD endpoint that matches the other endpoints, set up a new page from existing components, write tests for existing functions.
These are tasks where the “what” is fully specified and the “how” can be inferred from code that already exists. Hand it that setup and it delivers solid junior-developer output without the onboarding, the standups, or the sick days. If your team has a genuinely deep queue of that work, the autonomy is real.
Where It Falls Over
Devin struggles with anything requiring judgment: deep context about business logic, debugging interactions between systems, or matching a visual design. These are precisely the tasks where expertise and context matter more than code generation speed — which is to say, most of the interesting ones.
The feedback-loop problem
Here is the comparison nobody at Cognition wants made: Claude Code with an experienced developer was faster and produced better results than Devin working alone on every single task we tested. The difference is not model quality. It is the feedback loop.
With Claude Code, you course-correct in real time. “No, the filter belongs on the database query, not the application layer.” “Cover the rate limit case in the error handling.” Those corrections take seconds and prevent twenty minutes of confidently wrong autonomous work.
Devin's autonomy looks phenomenal in a demo. In practice, the lack of human steering means it commits early to a flawed approach and then iterates hard on the flaw. We have written before about why the human-in-the-loop model keeps winning — Devin is the strongest evidence yet.
Who Should Actually Buy It
At $500 per month, Devin is hard to justify unless you have a large volume of well-specified, pattern-matching tasks queued up. If your backlog is fifty near-identical CRUD tickets, sure — throw the robot at it and go do something more fun.
For most teams, Claude Code at $20 per month plus a human in the loop gets better results on everything, including the routine stuff. And no, this is not the moment juniors get replaced either — Devin competes for the same routine work and is currently losing it to cheaper tools. We covered what AI actually means for junior devs separately; short version, the panic is aimed at the wrong target.
The Verdict
Devin is not a replacement for developers. It is not even close. But it is also not useless — it is a semi-autonomous junior that handles routine work well, provided you hand it clear specifications and existing patterns to follow.
The hype was wrong, but the underlying technology will improve. The real question is whether Devin can improve fast enough to justify its positioning before Claude Code and Copilot Workspace close the gap from the other direction. Plot twist: they are closing it fast, and they cost a fraction as much.
At Aidxn Design we ship client builds with AI coding tools — Claude Code and Cursor — in the loop every day, with a senior human steering. That combination is the actual future, and it is how we deliver production web development that does not need a babysitter.