Walden Yan

Walden Yan is a co-founder of Cognition, the company behind Devin, an autonomous software-engineering agent. His work spans neural program synthesis and the design of agents that carry out extended engineering tasks. His writing explains a particular difficulty in that progression: agents can produce individually plausible changes that conflict when combined, because each change carries decisions about how the software should work.

From competitive programming to program synthesis

Walden taught himself programming through YouTube videos. Competitive programming developed both his familiarity with algorithms and his ability to tackle problems that resisted familiar techniques. As a first-year Harvard student in 2020, he earned a gold medal at the International Olympiad in Informatics, finishing 19th among nearly 400 competitors. He estimated that he had spent more than a thousand hours practicing. Afterward, he planned to step back from contests to explore his studies, including machine learning and website design.

Through MIT PRIMES, Walden collaborated with William S. Moses on token pairing for neural program synthesis. Their research addressed a structural mismatch: programs have nested relationships, while a sequence-to-sequence model generates code as a linear stream. They combined token pairing with attention and syntax guidance, training and evaluating the model on AlgoLisp problems that paired English descriptions with programs and test cases.

The project repository reports 99.24% evaluation accuracy, compared with a previous state-of-the-art result of 95.80%, while using fewer parameters. The collaborators also created an interpreter that corrected bugs in AlgoLisp’s built-in executor—important because evaluating generated code depends on executing it correctly. Walden’s repository contains the implementation, dataset preparation, and training and evaluation workflow; these results belong to the joint research.

At Harvard, Walden studied computer science and economics and explored cryptographic applications. In 2022, he extended private blockchain mixing tools using zero-knowledge proofs, working on software that allowed users to deposit, withdraw, and trade assets without making those transactions publicly traceable. His public projects also include EnhanceAI’s React wrapper, which integrates AI autocomplete into web application text inputs.

Letting an agent carry out the task

Walden recounts an early Devin experiment in which the team struggled to start a MongoDB server. He asked their rough prototype to get it working. It ran terminal commands and deleted files while the team watched with some concern; eventually, the server worked. For Walden, who had wanted to message an agent and have it build his projects, the experience raised a concrete question: how much further could they push it?

Cognition introduced Devin in March 2024, equipping it with a shell, editor, and browser in a sandboxed development environment. The company presented planning, implementation, debugging, and testing as parts of a continuing task, with progress reports and opportunities for user feedback. Walden’s later essays examine how agents should retain and exchange the information needed for that work.

Why shared instructions are not enough

In his June 2025 essay on context engineering, Walden argues that an agent needs the history of relevant conversations, tool results, and decisions. Giving several agents the same initial request does not ensure that their work will fit together: they continue making choices during execution.

His example is a game divided between agents building a bird and a background. Each can finish its assignment while choosing incompatible visual styles or behavior. The problem is not simply misunderstanding the request; neither agent can see the assumptions emerging from the other’s work. Walden therefore recommended a single continuing agent as the practical default. For longer tasks, he proposed compressing its history into consequential details, events, and decisions, while acknowledging that identifying what to preserve requires substantial engineering.

By April 2026, he had revised that recommendation around patterns Cognition had found useful: additional agents could contribute intelligence while writes stayed single-threaded.

• Fresh-context review: A reviewer examines the implementation without inheriting the coding agent’s accumulated assumptions and long working history. It can reason backward from the code and question choices the coder overlooked. The coding agent must then assess the findings against the user’s instructions and earlier decisions, so review does not create repeated loops or expand the task beyond its scope.

• Consulting another model: Cognition’s experiment pairing SWE-1.5 with a stronger adviser offered speed and cost advantages, but the weaker primary model struggled to recognize when it needed help and what to ask. Walden describes better results when both models were capable frontier models, with consultation directed toward their different strengths in debugging, visual reasoning, or testing. The cheaper-primary arrangement remained an open problem rather than a solved route to frontier performance.

This development preserves the reason for his earlier caution. He became more receptive to collaboration where communication and synthesis could be engineered, while remaining skeptical of parallel writers making conflicting decisions.

Engineering after deployment

Walden’s scope also includes keeping software running. In welcoming TierZero to Cognition in July 2026, he credited Anhang, Yun, and their team’s automation work and described plans to bring it into Devin. He emphasized catching incidents before they spread and maintaining system health over time, with the aim of reducing the time engineers spend handling incidents. This was an integration ambition, extending the agent’s responsibilities beyond producing and shipping code.