Cognition

Cognition builds Devin, an autonomous software engineering agent, and developer tools for understanding codebases and managing agent work. Its acquisition of Windsurf brought an IDE into the company’s product suite; that foundation became Devin Desktop in June 2026. Cognition’s strategy combines its own models with those of outside providers. In its September 2026 Series E announcement, the company reported raising over $2 billion at a $48 billion valuation and reaching almost $900 million in run-rate revenue.[3][8][11][12]

The company’s cloud-agent work preceded its expansion into IDEs. Cognition says it started in 2024; Scott Wu later recalled a small group of engineers with more than a decade of friendship building the early product in a New York apartment. Walden Yan, identified as Cognition’s cofounder and Chief Product Officer in May 2026, contributed its published work on agent architecture and context engineering.[12][8][1][6]

Wu introduced Devin on March 12, 2024. It combined planning and context recall with a shell, code editor and browser inside a sandboxed compute environment. Users could delegate tasks for review or collaborate through progress reports, feedback and design decisions. The launch announcement also disclosed a $21 million Series A led by Founders Fund.[3]

Launch demonstrations covered more than code generation: Devin read about ControlNet and ran it on Modal to produce images with concealed messages for Sara; built a Game of Life website, incorporated user requests and deployed it to Netlify; and helped Andrew debug his open-source competitive programming book. In a SymPy logarithm-bug example, it set up the environment, reproduced the bug, wrote the fix and tested it.[3]

Cognition reported that Devin resolved 13.86% of SWE-bench issues end to end at launch. The evaluation used a random 25% subset of the dataset. Devin was unassisted, while comparator models were told which files to edit; the result describes that historical evaluation rather than a general coding success rate.[3]

Devin 2.0, released April 3, 2025, added an interactive cloud IDE and a plan starting at $20. Each parallel Devin session had its own IDE, where users could inspect changes, edit code and run tests. Interactive Planning researched the codebase and presented findings and an editable preliminary plan before autonomous execution. Devin Search answered codebase questions with cited code, while Devin Wiki periodically generated documentation, architecture diagrams and source links.[4]

On May 5, 2025, Cognition launched DeepWiki, the free public version of Devin Wiki and Devin Search. Its announcement instructed users to replace github.com with deepwiki.com in a repository URL and said more than 50,000 top public GitHub repositories had already been indexed. Public repositories could be added free; private repositories required a Devin account.[5]

On July 14, 2025, Cognition signed a definitive agreement to acquire Windsurf, including its IDE, intellectual property, brand, business and personnel. The announcement reported Windsurf’s scale as $82 million ARR, more than 350 enterprise customers and hundreds of thousands of daily active users. Wu said he and Jeff structured the deal so every Windsurf employee would participate financially, have vesting cliffs waived and receive fully accelerated vesting for work to date.[7]

By September 2025, Wu described the acquisition as accomplished and explained the complementary product forms: an IDE lets engineers make individual decisions with AI assistance, while agents accept complete tasks delegated asynchronously. Cognition reported less than 5% overlap between the businesses’ enterprise customers before the acquisition. The Windsurf acquisition more than doubled Cognition’s ARR, and combined enterprise ARR grew over 30% in the seven weeks after the purchase.[8]

Scott Wu and Jeff Wang jointly announced Devin Desktop on June 2, 2026 as the next generation of Windsurf, following Windsurf 2.0’s Agent Command Center and Devin integration. Desktop made the command center the default IDE surface for managing local and cloud agents, pull requests and context. Spaces enabled related agents to share context, and Agent Client Protocol support allowed ACP-compatible agents to run alongside Devin. The full IDE retained editing and review workflows; the authors described it as backwards-compatible with Windsurf.[11]

A year before Desktop’s launch, Yan’s June 2025 essay addressed context management, a problem also central to coordinating agents in the later product. He argued against multi-agent systems that fragment context and decisions. His illustrative Flappy Bird example showed how separately assigned agents could produce incompatible backgrounds and bird behavior, even when given the same task. He recommended sharing full agent traces and accounting for the implicit choices carried by actions. A linear agent preserves continuous context; for longer tasks, Cognition had fine-tuned a smaller model to compress action and conversation history into important details, events and decisions.[6][11]

Yan’s April 2026 follow-up identified a narrower useful pattern: multiple agents contribute intelligence while writes remain single-threaded. Devin and Devin Review iterated on coding and review, with the reviewer starting from clean context rather than inheriting the coder’s reasoning. The coding agent then reconciled findings with its broader instructions and decisions to prevent loops and work outside the task. Yan reported an average of two bugs caught per Devin-written pull request, roughly 58% severe. Manager Devins could also split larger tasks, spawn child Devins and coordinate progress through an internal MCP, with context transfer and communication still requiring further work.[9]

The same follow-up preserved an unsuccessful experiment. Smart Friend let a primary model consult another model when needed, but SWE-1.5 was insufficiently capable at deciding when to escalate and what to ask Sonnet 4.5. SWE-1.6 improved the pattern without fully solving it. Yan reported better results from production consultation between Claude and GPT, where delegation could route a subtask to the model best suited to debugging, visual reasoning or testing.[9]

In a May 2026 interview, Yan described Devin’s architecture as separating “the brain from the machine”: the control system remained inaccessible from the working machine, which received narrowly scoped credentials. Full virtual machines supported running applications, Docker-based dependencies and recorded testing. Cognition also built incremental disk-saving infrastructure so restoring an agent’s machine could involve work proportional to filesystem changes; Yan noted that the described storage system had already been superseded internally. He described infrastructure, integrations and customer onboarding as parts of what Cognition supplied with Devin.[1]

Cognition extended its own model work with SWE-2 on September 10, 2026. Post-trained from the Kimi Team’s 2.8-trillion-parameter Kimi K3, it used a single reinforcement-learning run across reasoning-effort levels. Effort-specific linear cost penalties balanced successful completion against inference cost and rollout time; a length-weighted reward baseline sought to stabilize training. The rollout system used DSpark speculative decoding, SpecForge-trained draft models and online draft-model training to keep proposals aligned with the changing policy. Cognition reported 50.0% on FrontierCode 1.1 Main and 73.0% on DeepSWE 1.1, selecting each model’s best effort-setting score and combining public results with internal evaluations using different primary harnesses. SWE-2 launched in Devin Desktop and CLI, with Web and Fusion rolling out.[13]

Cognition’s September 8, 2025 financing announcement reported over $400 million raised at a $10.2 billion post-money valuation, led by Founders Fund. Wu reported standalone Devin ARR rising from $1 million in September 2024 to $73 million in June 2025, before Windsurf, with cumulative company net burn below $20 million. On May 27, 2026, Cognition announced over $1 billion raised at a $26 billion valuation, led by Lux Capital, General Catalyst and 8VC, and reported enterprise usage growth above tenfold since the start of the year and $492 million in run-rate revenue.[8][10]

The September 8, 2026 Series E was led by new investors Andreessen Horowitz and Accel alongside existing investors Founders Fund, General Catalyst and Avenir. Cognition reported run-rate revenue growing from May’s $492 million to almost $900 million; these figures were described as run-rate revenue rather than ARR or annual revenue. It also announced offices opened during the preceding year in Washington, D.C., Tokyo, Singapore, London, São Paulo and Madrid, alongside its San Francisco, New York and Austin hubs.[12]

By that announcement, Devin’s product scope included proactive work beyond individually prompted coding tasks. Auto-Triage took a first pass at incident investigation, Security Swarm found and triaged vulnerabilities, and Automations started work from events in Slack, GitHub, Linear and other systems. Cognition continued to describe itself as an independent agent lab able to choose and combine outside models and its own.[12]