Trustvice 360
Somewhere inside a large bank there is a codebase older than some of the engineers who look after it. It runs to millions of lines and thousands of files, much of it written by people who left years ago. Nobody alive knows all of it. Every change starts with days of archaeology, and the big jobs, a migration to a new language or a new version of a framework, sit on the roadmap for years because there are never enough people to do them.
That is the kind of place Devin was made for. It reads the code, writes down how the system fits together, and then does the work: one ticket, one migration, one pull request at a time, often many of them at once.
From a demo to a colleague#
Cognition introduced Devin on March 12, 2024, as the first AI software engineer. The company was founded by Scott Wu, Steven Hao and Walden Yan, three champion competitive programmers, and its launch video showed something new: an AI with its own shell, code editor and web browser, planning a task, writing the code, hitting an error, reading the logs and fixing it. On SWE-bench, a test built from real bugs in open-source projects, it solved 13.86 percent of a sample of issues. The best result before it had been 1.96 percent.
Devin became generally available in December 2024. Devin 2.0 followed in April 2025 with a cloud development environment, several Devins working in parallel and plans from $20 a month. In July 2025 Cognition acquired Windsurf, the AI code editor, along with its team and its enterprise customers, and in June 2026 Windsurf's editor returned as Devin Desktop.
Growth since then has been steep. Cognition says its annualized revenue run rate passed $1 billion in September 2026, the same month a Series E led by Andreessen Horowitz and Accel valued the company at $48 billion.
A computer of its own#
The idea at the heart of Devin is simple: give the AI the same tools a human engineer has, and let it use them on its own.
Every Devin session runs on its own virtual machine with a terminal, an editor and a browser. Devin can install dependencies, run the test suite, start the app and click through it. It now has a full Linux desktop, records its screen while it tests, and can check its work on Windows and Mac machines too.
People hand Devin work wherever they already talk about work: tag it in Slack or Microsoft Teams, assign it a Linear or Jira ticket, point it at a failing GitHub check, or type into the web app. It writes a plan, asks questions when something is unclear, and comes back with a pull request that has been built and tested. Because it runs in the cloud, it keeps working when the laptop is closed.
For bigger jobs, one Devin can manage others. A coordinator splits the work, hands each piece to a child Devin running in its own machine, and keeps track of the whole. Automations start sessions by themselves when an event arrives, such as a new ticket or a broken build. Developers who prefer their own machine can start in the Devin command-line tool or in Devin Desktop and hand a task off to the cloud when it grows.
Maps for the maze#
What sets Devin apart in a very large codebase is how much effort goes into understanding the code before changing it.
DeepWiki turns a repository into a wiki: an overview of the system, architecture diagrams and pages for each part, every claim linked to the source it came from. For open-source projects it is free for anyone. Take a GitHub address, replace github.com with deepwiki.com, and a guided tour of the project appears. It launched in May 2025 with more than 50,000 popular public repositories already indexed. For a company's own code, DeepWiki indexes every connected repository and keeps the pages fresh as the code changes.
Ask Devin answers questions about the code in plain language, with citations to the exact files and lines. Ask where a payment is retried or which services read a certain table, and it shows its sources. When the answer turns into a task, Ask Devin can write the plan and start a session with all that context already attached.
Devin also learns the house rules. It picks up knowledge from the notes teams already keep, such as READMEs and the instruction files many projects now write for AI agents, and it builds reusable instructions for recurring work. Environments are set up once as blueprints, so every session starts from a fresh, fully built copy of the system, with several repositories checked out side by side if a change spans them.
The scale is the point. Cognition says DeepWiki has documented five million lines of COBOL, the decades-old language still running much of the world's banking, as well as repositories of 500 gigabytes, and that one bank generated documentation for more than 400,000 repositories.
The jobs nobody wanted#
Large organizations carry work that is important, enormous and dull: moving millions of lines to a new framework, upgrading a language version across hundreds of services, fixing every finding from a security scanner. It is exactly the work Devin's design suits, because it can be split into many similar pieces and checked by tests.
The customer stories Cognition publishes read like a list of those jobs:
- Nubank, the Brazilian digital bank, had an eight-year-old data-processing monolith of more than six million lines, a migration once estimated at 18 months and over a thousand engineers. With Devin, Cognition reports, the work took a fraction of the engineering hours, eight to twelve times more efficient, at more than twenty times lower cost.
- Itaú, Latin America's largest bank, used Devin to document more than 300,000 repositories. Cognition says about 70 percent of findings from its security scanners are now fixed automatically, test coverage rose from about half the code to more than 90 percent, and a migration of 59 services from .NET to Java ran six times faster at a fifth of the cost.
- Mercedes-Benz had Devin analyze more than 200,000 lines of COBOL, a job estimated at eight months that took eight days.
- Ramp, the finance platform, reports more than 10,000 hours saved a month and around 80 merged pull requests a week.
The biggest banks have moved too. Goldman Sachs began piloting Devin alongside its roughly 12,000 developers in 2025, and Citi started a rollout to about 40,000 developers the same year, with people, not agents, deciding what gets deployed.
How it differs from other AI tools#
Most AI coding tools sit next to a developer. Assistants such as GitHub Copilot, and editors such as Cursor, suggest and write code while a person types. Terminal agents such as Claude Code and OpenAI's Codex CLI work on the developer's own machine, inside the task the developer is doing right now. Cloud agents from GitHub and OpenAI can take an issue and return a pull request.
Devin is built around delegation. Cognition describes it as an autonomous engineer rather than an editor. A few things follow from that:
- Its own machine for every task, so it can build, run and test the whole system rather than suggest lines in a file.
- Memory across a whole organization: DeepWiki, knowledge and past sessions span every connected repository, not one open project.
- Work in parallel: many Devins, coordinated by a Devin, each in its own machine.
- Enterprise deployment: options range from a shared cloud to a single-tenant private network with customer-managed keys, and Devin Desktop and the command-line tool are authorized at FedRAMP High for US government work.
Devin is not tied to one AI model either. Cognition trains its own coding models, the SWE series, and its Fusion approach pairs a frontier model in the lead with a cheaper partner, which the company says cuts costs by up to 60 percent. Devin Desktop can even run other agents, such as OpenAI's Codex and Anthropic's Claude, side by side on one board.
Review, security and a guarantee#
Writing code is half the job. Devin Review reads a pull request the way a careful colleague would: it groups and orders the changes, recognizes code that was only moved or renamed, and marks issues by how much they matter. It is free for open-source projects. On the security side, Devin can sweep a codebase for vulnerabilities and fix what it finds, and a deeper security scan arrived in October 2026.
Cognition is confident enough to put money behind it. Under its AI Productivity Guarantee for enterprise customers, an estimator puts a value in engineer-hours on each session. If that value falls short of what the customer spent by the end of the contract, Cognition covers the difference in usage credits, up to $10 million.
Getting started is cheap. There is a free plan, Pro costs $20 a month and Max $200, Teams starts at $80 a month plus $40 per developer seat, and Enterprise is priced to fit.
A teammate that never forgets the codebase#
The promise of Devin is not that software engineers disappear. It is that the backlog does: the migrations postponed for years, the documentation nobody had time to write, the security findings that pile up faster than people can fix them. Engineers decide what should be built, and a patient colleague with a perfect memory of the codebase does more and more of the building.
The easiest way to feel it costs nothing. Pick a favorite open-source project on GitHub, swap github.com for deepwiki.com in the address, and read the map. Then imagine one like it for every codebase you have ever been afraid to touch. The details live at devin.ai and in Devin's documentation.
