Walden Yan

Walden Yan is a co-founder of Cognition and was its chief product officer as of May 2026. His work on Devin, an AI software engineering agent, includes technical arguments about maintaining context across long tasks and product decisions about how agents run, test applications, and respond to people. Between June 2025 and April 2026, he moved from warning against multi-agent architectures to describing narrower forms of collaboration in which additional agents contribute intelligence while writes remain single-threaded.[1][2][3]

In his June 12, 2025 essay Don’t Build Multi-Agents, Yan described context engineering as automatically supplying models with the information they need within a dynamic system. He emphasized two principles: share full agent traces rather than isolated messages, and recognize that actions contain implicit decisions. Giving parallel agents the same task does not ensure that they make compatible choices about style, implementation, or edge cases. Yan later acknowledged that others had used the term context engineering before him.[2][1]

Yan illustrated the coordination problem with a hypothetical Flappy Bird clone: one agent builds the background and another builds the bird, but their interpretations or visual styles diverge. The agent combining their work inherits those conflicts. His simplest recommendation was a single-threaded linear agent with continuous context. For longer tasks, he proposed a model that compresses action and conversation history into important details, events, and decisions, reporting that Cognition had fine-tuned a smaller model for this purpose. He stressed that compression requires careful selection of information and still eventually reaches limits.[2]

In Multi-Agents: What’s Actually Working, published April 22, 2026, Yan retained his objections to parallel-writer swarms but described practical alternatives. His organizing distinction was between agents contributing intelligence and agents independently taking conflicting write actions.[3]

One alternative was a coding agent paired with a reviewer. Yan reported that this loop worked best when the reviewer initially received a clean context rather than the coder’s accumulated instructions and reasoning. The reviewer could inspect the diff, reconstruct relevant context, and question assumptions the coder had accepted. The coding agent then needed to assess the findings against the broader task and user decisions. That communication bridge, he argued, helped prevent repeated loops, unnecessary work, and changes outside the requested scope.[3]

Another experiment, Smart Friend, let a primary model consult a second model. Yan described an important failure: SWE 1.5 was not sufficiently capable of recognizing when to escalate or knowing what to ask Sonnet 4.5. Its speed and cost advantages did not overcome that quality ceiling. He reported that SWE 1.6 improved the pattern but still fell short of the team’s goal. Separately, he reported production gains from pairing Claude and GPT, where delegation routed work to differing capabilities—such as debugging, visual reasoning, or writing tests—rather than simply escalating to a stronger model.[3]

Yan also described manager Devins dividing larger tasks, spawning child Devins, and coordinating through an internal MCP. He favored “map-reduce-and-manage”: a manager splits work, children execute, and the manager synthesizes the results. Making this coherent required addressing overly prescriptive managers, mistaken assumptions about shared state, and children failing to communicate discoveries that should affect other agents’ work. These remained active engineering problems in his April account.[3]

In a May 28, 2026 interview, Yan explained Cognition’s decision to separate Devin’s “brain” from its development machine. This allowed the team to reuse development infrastructure and place narrowly scoped credentials on the machine while keeping the brain inaccessible from it. He connected that separation to permission management, including cases where different users interact through the same GitHub application.[1]

Yan explained the team’s choice of full virtual machines through the need to run and test real applications, including applications that themselves use Docker. He treated testing as a broader reasoning problem than clicking interface coordinates: an agent must start cooperating services with the right code versions, arrange permissions or feature flags, and trigger the behavior being tested. He reported cases in which Cognition had to orchestrate multiple frontier models to complete that work. He also said he had added annotations to testing videos so viewers could understand what Devin was demonstrating.[1]

Yan described the team tuning Devin’s GitHub interactions so it could receive review comments, address useful findings, and push back when a comment was wrong. Avoiding unproductive review loops required judgment about which comments to act on. Slack posed a related product problem: giving an agent tools to post messages was insufficient for a coworker-like experience. Cognition built incoming webhook support and tuned conversational responses, including how to avoid spamming threads.[1]

Yan credited Cognition’s infrastructure colleagues with diagnosing slow grep searches as network-backed filesystem overhead; the team replaced that filesystem. He distinguished this fix from the separate BlockDev storage format, which reduced VM save-and-restore work by processing filesystem changes rather than the entire disk. By the May interview, he said that described storage system had already been superseded internally.[1]

Yan also recounted internal experiments building intended products through successive AI changes with automatic merging and no code review. He estimated that the December state of the art sustained this approach for about two weeks before maintenance became troublesome. His example was a button implemented in ten places with inconsistent variations, making a color change difficult. The team then introduced code review and cleanup. The duration was his account of those experiments, rather than a general limit on AI-written software.[1]

Before these essays and interviews, Yan was already working on Cognition while associated with Harvard. A March 27, 2024 Harvard Crimson report identified him as a Harvard student in the class of 2024, Cognition AI’s co-founder and chief product officer, and a member of the 2024 Thiel Fellowship class. The fellowship awarded $100,000 grants to pursue startups and required recipients to leave school.[4]