An engineer can spend the morning writing a patch and the afternoon waiting for someone to approve it. For an AI engineer, the waiting is a surprisingly awkward problem. Where does the half-finished work live? Who pays for the machine while it sits idle? And when a reviewer finally says, “Change this,” can the agent remember what it was doing?
Cognition has spent much of its short life working on questions like these. Its product, Devin, takes an assignment, works inside a development environment, tests changes and returns code for review. The ambition is delegation: give the machine a piece of the backlog and get on with another problem.
- The product: a cloud coding agent, an editor and tools for understanding and reviewing code.
- The customers: developers and engineering organizations, from startups to banks and automakers.
- The business: subscriptions, paid usage and enterprise contracts.
- The useful lesson: clear assignments and rigorous review matter more as execution gets cheaper.
The engineer who can wait
On September 25, 2026, Cognition reported crossing $1 billion in annualized revenue run rate. That measures the business’s current revenue pace projected over a year; it does not mean the company has booked a billion dollars over the preceding twelve months. Still, it is a considerable audience for a product introduced in March 2024.
The work itself is often unromantic. A library needs upgrading. A test is missing. A framework migration has been postponed again. Engineers at Citi, GE Aerospace, NVIDIA and Modal are among the customers Cognition names. These organizations have existing systems, established conventions and very little patience for a beautiful demonstration that fails in their repository.
Devin’s cloud infrastructure is part of its answer. Cognition says it invested more than a year in microVM engineering, isolating sessions and preserving entire machine states. An agent can stop consuming active compute while waiting, then resume with its memory, processes and files intact. The humble pause button turns out to require quite a lot of engineering.
Ten gold medals, one apartment
The founders are Scott Wu, the CEO; Steven Hao, the CTO; and Walden Yan, the chief product officer. The founding team collectively held ten gold medals from the International Olympiad in Informatics. Cognition also says 21 of its first 35 team members had already founded a company. This was an unusually experienced group for something that began in a New York apartment.
Programming contests reward the ability to choose a route through a difficult problem under pressure. Building an agent adds another demand: turn that judgment into a system that can recover when a command fails or a test exposes an error. Devin’s launch paired language-model reasoning with a shell, an editor and a browser. It could act on its answers and inspect the consequences.
There is a distinction between winning a contest and maintaining someone else’s software. The second requires fitting into the house style, preserving behavior and knowing when to ask. Cognition’s trajectory is interesting because it has had to build for that second examination.
The weekend that bought the other half
In July 2025, Windsurf’s leaders and some researchers left for Google. The remaining business still had an AI editor, customers and a team. Cognition moved quickly: the companies describe a first call on Friday evening and a signed agreement by Monday morning.
The fit was concrete. Devin handled delegated cloud work; Windsurf provided the place where engineers worked interactively. Windsurf also brought a sales organization and, at the announcement, $82 million in annual recurring revenue. Cognition’s September retrospective said fewer than 5% of their enterprise customers overlapped. Buying the editor widened both the product range and the customer base.
“there’s only one boat and we’re all in it together.”
Scott Wu and Jeff Wang
Recalling the Windsurf integration, July 2026
That combination became Devin Desktop in June 2026. It retains a full IDE, but makes an Agent Command Center the default surface. Spaces let related agents share context; support for the Agent Client Protocol allows compatible outside agents to join. Engineers can delegate, monitor and then step back into the code themselves.
This puts Cognition in competition with Cursor, GitHub Copilot, Claude Code and Codex across overlapping parts of the development workflow. Its proposition bundles asynchronous execution, interactive editing and the infrastructure around them. Cognition also uses outside models alongside its own, selling independence from any single model supplier.
The bill arrives in tokens
Devin’s December 2024 general release started at $500 a month. By October 2026, the pricing page lists Free, Pro at $20 monthly and Max at $200. Teams carries an $80 monthly plan fee plus $40 per full developer seat; enterprise contracts have custom terms. Paid plans include usage allowances, with extra consumption available at API pricing.
The entry price is only the beginning of the calculation. Task complexity, model choice and reasoning affect consumption. More sessions can produce more useful work, but they can also produce more work for reviewers. A sensible trial counts finished assignments, total spend and the human effort needed to get a change accepted.
Cognition’s April pricing explanation was unusually candid: the old team entry point deterred adoption, while newer tools consumed meaningful compute. Its business model has to reconcile an inviting front door with an expensive kitchen.
Several brains. One pair of hands.
Walden Yan published advice against building multi-agent systems in 2025. In April 2026, he revisited it. The original difficulty was coordination: parallel agents quietly chose different styles, assumptions and edge-case behavior. Their individually reasonable edits could leave a collectively fragile product.
What changed his position was a narrower pattern: several agents could contribute intelligence while edits remained under coherent control. Better models and heavier customer usage also made planning and review more pressing. One experiment paired a coding agent with a separate reviewer that began with clean context, allowing it to question the implementation afresh.
- 01ScopeDefine a checkable outcome.
- 02ImplementAgent edits and tests.
- 03ReviewA fresh pass finds issues.
- 04DecideHuman approves the change.
The lesson is easy to copy: give review its own attention, and keep responsibility for changes explicit. A crowd of clever assistants still needs a clear account of what the job is. Buying more intelligence does not purchase agreement.
A patch has to earn its place
In November 2025, Cognition reported that Devin’s pull-request merge rate had risen from 34% to 67% over a year. The same assessment described difficulty with ambiguous requirements and mid-task scope changes. It recommended clear, verifiable assignments and retained human review. Those are historical findings, but useful questions for a present-day pilot.
FrontierCode, introduced in June 2026, extends that concern into evaluation. Its criteria include correctness, test quality, scope and conformity with a codebase’s standards. Passing tests is one hurdle. Persuading a maintainer to accept the patch is the larger business problem.
Cognition’s enterprise productivity guarantee takes a related approach. It estimates useful output in human-equivalent hours, converts those hours into value and compares that with consumption. Under the announced terms, shortfalls can receive usage credits up to $10 million. The estimate is a baseline for productivity; it does not establish the business value of every task.
Give it a job you can judge
A practical first assignment might be one dependency upgrade with a documented test command and an explicit boundary on changed files. Ask for a plan, inspect the result, record the cost and repeat on comparable work. This makes the trial informative even when an agent falls short.
Enterprise adoption adds permissions, network access and deployment rules. Cognition’s September 2026 AWS collaboration addresses modernization and backlog work on infrastructure customers already use. Its dedicated deployments and account engineering matter because a capable agent becomes useful only when it can reach the necessary systems.
Broad briefs, weak tests and an overwhelmed review queue are poor conditions for delegation. Increasing execution capacity can amplify those weaknesses. Cognition’s mission is to give engineers more room to act as architects. The price of that freedom is a more exacting brief, a sharper review process and the discipline to merge only work that deserves a place.