# Walden Yan

![Walden Yan](/resources/cognition/portraits/walden-yan.webp)

Human @ Cognition AI | IOI ’20 Gold Medalist | MIT PRIMES Researcher  
Cognition  
San Francisco, California, United States

## At Cognition

- Role: Co-Founder & CPO (Leadership)
- Joined: Nov 2023
- How they joined: Directly
- Before Cognition: Early Engineer, Cursor
- Olympiads: IOI gold 2020 (https://www.autumn.ai/resources/olympiads/walden-yan-olp-4ba8fd82b68c5f8b813f)
- Overlapped before Cognition with other Cognition people at: Harvard University (https://www.autumn.ai/resources/cognition/worked-together)

## Identity
- LinkedIn: https://www.linkedin.com/in/waldenyan
- X: @walden_yan
- GitHub: https://github.com/walnutwaldo
- Website: https://waldenyan.com
- Domain: cognition.ai

## Notes
- In a May 28, 2026 interview, Walden Yan explained Cognition's decision to separate Devin's "brain" from its development machine. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden 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. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan treated testing as a broader reasoning problem than clicking interface coordinates. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan reported cases in which Cognition had to orchestrate multiple frontier models to complete testing work. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan said he had added annotations to testing videos so viewers could understand what Devin was demonstrating. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden 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. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan credited Cognition's infrastructure colleagues with diagnosing slow grep searches as network-backed filesystem overhead, and the team replaced that filesystem. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan distinguished the grep filesystem fix from the separate BlockDev storage format, which reduced VM save-and-restore work by processing filesystem changes rather than the entire disk. (https://latent.space/wiki/walden-yan-7c572c23/)
- By the May interview, Walden Yan said the described storage system had already been superseded internally. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan recounted internal experiments building intended products through successive AI changes with automatic merging and no code review. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan estimated that the December state of the art sustained this approach for about two weeks before maintenance became troublesome. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan's example was a button implemented in ten places with inconsistent variations, making a color change difficult. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan was Cognition's chief product officer as of May 2026. (https://latent.space/wiki/walden-yan-7c572c23/)
- In Multi-Agents: What's Actually Working, published April 22, 2026, Walden Yan retained his objections to parallel-writer swarms but described practical alternatives. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan's organizing distinction was between agents contributing intelligence and agents independently taking conflicting write actions. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan described a coding agent paired with a reviewer as one alternative, reporting that this loop worked best when the reviewer initially received a clean context rather than the coder's accumulated instructions and reasoning. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan described an experiment called Smart Friend that let a primary model consult a second model. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan described an important failure in Smart Friend: SWE 1.5 was not sufficiently capable of recognizing when to escalate or knowing what to ask Sonnet 4.5. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan reported that SWE 1.6 improved the Smart Friend pattern but still fell short of the team's goal. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan 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. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan described manager Devins dividing larger tasks, spawning child Devins, and coordinating through an internal MCP. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan favored "map-reduce-and-manage": a manager splits work, children execute, and the manager synthesizes the results. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan works at Cognition, where agent building is a key frontier the company thinks about. Quote: "At Cognition, agent building is a key frontier we think about." (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan states that Cognition has fine-tuned a smaller model for compressing action and conversation history into key details, events, and decisions. Quote: "Depending on the domain, you might even consider fine-tuning a smaller model (this is in fact something we’ve done at Cognition)." (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan notes that Devin, among coding agents, IDEs, and app builders, used an 'edit apply model' in 2024. Quote: "A common practice among coding agents, IDEs, app builders, etc. (including Devin) was to use an “edit apply model.”" (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan wrote an article titled 'Principles of Context Engineering' published on 2025-06-12. (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan defines 'context engineering' as the next level beyond prompt engineering, about doing this automatically in a dynamic system, and calls it effectively the #1 job of engineers building AI agents. Quote: "“Context engineering” is the next level of this. It is about doing this automatically in a dynamic system. It takes more nuance and is effectively the #1 job of engineers building AI agents." (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan states his first principle of context engineering: share context, and share full agent traces, not just individual messages. Quote: "Share context, and share full agent traces, not just individual messages" (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan states his second principle of context engineering: actions carry implicit decisions, and conflicting decisions carry bad results. Quote: "Actions carry implicit decisions, and conflicting decisions carry bad results" (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan believes that libraries such as OpenAI's swarm and Microsoft's autogen actively push concepts he believes to be the wrong way of building agents, namely multi-agent architectures. Quote: "In some cases, libraries such as https://github.com/openai/swarm by OpenAI and https://github.com/microsoft/autogen by Microsoft actively push concepts which I believe to be the wrong way of building agents. Namely, using multi-agent architectures, and I’ll explain why." (https://cognition.com/blog/dont-build-multi-agents, https://github.com/openai/swarm, https://github.com/microsoft/autogen)
- Walden Yan argues that Principles 1 and 2 are so critical and so rarely worth violating that you should by default rule out any agent architectures that don't abide by them. Quote: "I would argue that Principles 1 & 2 are so critical, and so rarely worth violating, that you should by default rule out any agent architectures that don’t abide by them." (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan says the simplest way to follow the principles is to just use a single-threaded linear agent. Quote: "The simplest way to follow the principles is to just use a single-threaded linear agent" (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan is optimistic about the long-term possibilities of agents collaborating with one another but says that in 2025 running multiple agents in collaboration only results in fragile systems. Quote: "While I’m optimistic about the long-term possibilities of agents collaborating with one another, it is evident that in 2025, running multiple agents in collaboration only results in fragile systems." (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan personally thinks cross-agent context-passing will come for free as single-threaded agents get better at communicating with humans. Quote: "I personally think it will come for free as we make our single-threaded agents even better at communicating with humans." (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan says Cognition builds its internal tools and frameworks around these principles as a way to enforce these ideas. Quote: "We build our internal tools and frameworks around these principles we repeatedly find ourselves relearning as a way to enforce these ideas." (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan invites readers to try Cognition's work at app.devin.ai. Quote: "We welcome you to try our work at app.devin.ai." (https://cognition.com/blog/dont-build-multi-agents)
- Walden Yan encourages people who would enjoy discovering agent-building principles with Cognition to reach out to him. Quote: "And if you would enjoy discovering some of these agent-building principles with us, reach out to " (https://cognition.com/blog/dont-build-multi-agents)
- In his June 12, 2025 essay Don't Build Multi-Agents, Walden Yan described context engineering as automatically supplying models with the information they need within a dynamic system. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan emphasized two principles: share full agent traces rather than isolated messages, and recognize that actions contain implicit decisions. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan illustrated the coordination problem with a hypothetical Flappy Bird clone in which one agent builds the background and another builds the bird, but their interpretations or visual styles diverge. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan's simplest recommendation was a single-threaded linear agent with continuous context. (https://latent.space/wiki/walden-yan-7c572c23/)
- For longer tasks, Walden Yan 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. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan stressed that compression requires careful selection of information and still eventually reaches limits. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan tweeted that he sees a lot of people make the same mistakes building agents and that Cognition shared a few of the principles they use. Quote: "I see a lot of people make the same mistakes building agents. So we shared a few of the principles we use" (https://latent.space/p/cognition)
- Walden Yan's tweet linked to Cognition's blog post 'Don't Build Multi-Agents'. Quote: "https://cognition.ai/blog/dont-build-multi-agents…" (https://latent.space/p/cognition, https://cognition.ai/blog/dont-build-multi-agents…)
- A March 27, 2024 Harvard Crimson report identified Walden Yan 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. (https://latent.space/wiki/walden-yan-7c572c23/)
- The Thiel Fellowship awarded $100,000 grants to pursue startups and required recipients to leave school. (https://latent.space/wiki/walden-yan-7c572c23/)
- Public headline says "Human @ Cognition AI | IOI ’20 Gold Medalist | MIT PRIMES Researcher". (https://www.linkedin.com/in/waldenyan)
- Walden Yan is a co-founder of Cognition. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan worked on Devin, an AI software engineering agent. (https://latent.space/wiki/walden-yan-7c572c23/)
- Between June 2025 and April 2026, Walden Yan moved from warning against multi-agent architectures to describing narrower forms of collaboration in which additional agents contribute intelligence while writes remain single-threaded. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan later acknowledged that others had used the term context engineering before him. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan was named to the 2024 Thiel Fellowship class. (https://latent.space/wiki/walden-yan-7c572c23/)
- Walden Yan is the coiner of context engineering. Quote: "Walden Yan, coiner of context engineering and Chief Product Officer/Cofounder of Cognition, invited OpenInspect’s Cole Murray to talk about why the Devin is in the Details." (https://latent.space/p/cognition)
- Walden Yan is Chief Product Officer and cofounder of Cognition. Quote: "Walden Yan, coiner of context engineering and Chief Product Officer/Cofounder of Cognition, invited OpenInspect’s Cole Murray to talk about why the Devin is in the Details." (https://latent.space/p/cognition)
- Walden Yan is a Cognition co-founder and CPO. Quote: "In this episode, Cognition co-founder and CPO Walden Yan joins swyx alongside Cole Murray, creator of OpenInspect, to unpack why everyone is building their own Devin, what changed after the December 2025 model inflection, and why “spec to pull request” is now becoming a real production workflow." (https://latent.space/p/cognition)
- Walden Yan went from coining context engineering to building the infrastructure behind Devin's 7x PR growth. Quote: "From coining “context engineering” to building the infrastructure behind Devin’s 7x PR growth and jump from 16% to 80% of commits across Cognition repos, Walden Yan has had a front-row seat to the background-agent shift." (https://latent.space/p/cognition)
- Walden Yan built infrastructure behind Devin's jump from 16% to 80% of commits across Cognition repos. Quote: "From coining “context engineering” to building the infrastructure behind Devin’s 7x PR growth and jump from 16% to 80% of commits across Cognition repos, Walden Yan has had a front-row seat to the background-agent shift." (https://latent.space/p/cognition)
- Walden Yan out-coded hundreds of opponents and took home a gold medal in the 32nd International Olympiad in Informatics. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan was a first-year student at Harvard University at the time of the article. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan placed 19th out of nearly 400 competitors representing more than 100 nations at the 32nd International Olympiad in Informatics. (https://seas.harvard.edu/news/2020/10/coding-gold)
- The 32nd International Olympiad in Informatics was held remotely in 2020 and challenged students to solve three computer science problems per day with a five-hour time limit. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan said that at the end of the first day he was tied with two other people for 28th place, the last spot for a gold medal. Quote: "At the end of the first day I was right on the border. I was tied with two other people for 28th place, and that is last spot where you get a gold medal. So on day two, a lot was on the line." (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan said it ended up being pretty close and that it was super-gratifying that all his work paid off. Quote: "It ended up being pretty close. It is super-gratifying that all my work paid off and I accomplished what I wanted to accomplish." (https://seas.harvard.edu/news/2020/10/coding-gold)
- One graph theory problem at the competition required students to create a network with a specific number of unique paths between each pair of nodes. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Due to the remote format, competitors connected to virtual machines set up in Singapore that blocked all other inputs, cameras recorded students during their five-hour sessions, and a local proctor was required. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Chris Tanner, Lecturer in Computational Science, served as proctor for Walden Yan. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan estimated he spent over 1,000 hours practicing competitive programming. Quote: "If I had to estimate, I probably spent over 1,000 hours practicing. Between working through online databases of old problems, practicing different techniques, and learning about new algorithms, it is a very big time dedication." (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan said you can't stop programming for a few months because you'll be rusty, comparing it to a real sport where you have to keep training. Quote: "You can’t stop programming for a few months, because by the time you come back to it, you’ll be rusty. It is almost like a real sport in a sense that you have to keep training." (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan said training helped him recognize patterns, so that just reading a problem's introduction would set off light bulbs and he'd realize which algorithms to use by the end of the description. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan encountered one problem with a very odd, circular setup whose description was so strange he had trouble wrapping his mind around it. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan said programming lets you create so many complex, unique things just with your keyboard and screen, and that competitive programming is a very creative, rewarding process. Quote: "One of the things about programming is the fact that you can create so many complex, unique things just with your keyboard and your screen. Competitive programming throws you a bunch of interesting problems to think about. You get to try your hand at so many different things, so it is a very creative, rewarding process." (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan taught himself to program using YouTube videos. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan was looking forward to diving deeper into computer science at the John A. Paulson School of Engineering and Applied Sciences and was especially excited to learn more about machine learning and website design. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan planned to take a break from competitive programming to dig into his studies and find a career path that speaks to him. (https://seas.harvard.edu/news/2020/10/coding-gold)
- Walden Yan won a gold medal at IOI 2020, ranking 19th. (https://cphof.org/profile/codeforces:walnutwaldo20)
- Walden Yan scored 505.38/600 at IOI 2020. (https://cphof.org/profile/codeforces:walnutwaldo20)
- Walden Yan's Codeforces profile is walnutwaldo20. (https://cphof.org/profile/codeforces:walnutwaldo20)
- Walden Yan has an IOI profile at stats.ioinformatics.org/people/7322. (https://cphof.org/profile/codeforces:walnutwaldo20)
- Walden Yan is known on Codeforces by the handle walnutwaldo20. (https://codeforces.com/profile/walnutwaldo20)
- Walden Yan's Codeforces profile lists his location as Avon, United States. (https://codeforces.com/profile/walnutwaldo20)
- Walden Yan's Codeforces profile lists him as being from USACO. (https://codeforces.com/profile/walnutwaldo20)
- Walden Yan's Codeforces contest rating is 2247. (https://codeforces.com/profile/walnutwaldo20)
- Walden Yan's maximum Codeforces rank is grandmaster with a maximum rating of 2497. (https://codeforces.com/profile/walnutwaldo20)
- Walden Yan's Codeforces title is Master. (https://codeforces.com/profile/walnutwaldo20)
- Walden Yan has solved 777 problems on Codeforces for all time. (https://codeforces.com/profile/walnutwaldo20)
- Walden Yan is a founder of Cognition AI, Inc. Quote: "Founders - Scott Wu - Steven Hao - Walden Yan" (https://en.wikipedia.org/wiki/Cognition_AI)
- Walden Yan serves as CPO (Chief Product Officer) at Cognition AI. Quote: "Keypeople - Scott Wu (CEO) - Steven Hao (CTO) - Walden Yan (CPO)" (https://en.wikipedia.org/wiki/Cognition_AI)
- Cognition was founded in August 2023 by Scott Wu, Steven Hao, and Walden Yan. Quote: "Cognition was founded in August 2023 by Scott Wu, Steven Hao, and Walden Yan." (https://en.wikipedia.org/wiki/Cognition_AI)
- Walden Yan was a competitive programmer who won a gold medal at the International Olympiad in Informatics (IOI). Quote: "All three were competitive programmers who won gold medals at the International Olympiad in Informatics (IOI)." (https://en.wikipedia.org/wiki/Cognition_AI)
- Cognition's products include Devin AI, DeepWiki, and Devin Desktop. Quote: "Products - Devin AI - DeepWiki - Devin Desktop" (https://en.wikipedia.org/wiki/Cognition_AI)
- In March 2024, Cognition released a demo of its AI coding tool, Devin AI, which was said to be able to perform the tasks of a software engineer. Quote: "In March 2024, Cognition released a demo of its AI coding tool, Devin AI, which was said to be able to perform the tasks of a software engineer." (https://en.wikipedia.org/wiki/Cognition_AI)
- Cognition boasted that its 10-person team as of March 2024 had won a total of 10 IOI gold medals. Quote: "Cognition boasted that its 10-person team as of March 2024 has won a total of 10 IOI gold medals." (https://en.wikipedia.org/wiki/Cognition_AI)
- In July 2025, Cognition signed a definitive agreement to acquire Windsurf, an agentic IDE, after the fallout from Google's $2.4 billion acquihire of their CEO and other senior employees. Quote: "In July 2025, Cognition signed a definitive agreement to acquire Windsurf, an agentic IDE, after the fallout from Google's $2.4 billion acquihire of their CEO and other senior employees." (https://en.wikipedia.org/wiki/Cognition_AI)
- In June 2026, Cognition announced rebranding of Windsurf to Devin Desktop. Quote: "In June 2026, Cognition announced rebranding of Windsurf to Devin Desktop." (https://en.wikipedia.org/wiki/Cognition_AI)
- In September 2026, Cognition engineer Eric Lu announced the factorization of RSA-260, the largest RSA challenge number publicly factored, using a GPU implementation developed and operated with Devin. Quote: "In September 2026, Cognition engineer Eric Lu announced the factorization of RSA-260, the largest RSA challenge number publicly factored, using a GPU implementation developed and operated with Devin." (https://en.wikipedia.org/wiki/Cognition_AI)
- Walden Yan found himself tied at the edge of a gold medal on the first day of the 2020 International Olympiad in Informatics. (https://yespress.io/walden-yan)
- The 2020 International Olympiad in Informatics moved online, with students working through virtual machines hosted in Singapore, watched by cameras and local proctors, with five hours to solve each day's set of problems. (https://yespress.io/walden-yan)
- Walden Yan was a first-year Harvard student at the time of the 2020 International Olympiad in Informatics. (https://yespress.io/walden-yan)
- Walden Yan finished 19th among nearly 400 competitors from more than 100 nations at the 2020 International Olympiad in Informatics, earning gold. (https://yespress.io/walden-yan)
- Walden Yan learned to code with YouTube videos. (https://yespress.io/walden-yan)
- Walden Yan estimated he had more than 1,000 practice hours before the IOI. (https://yespress.io/walden-yan)
- Walden Yan compared competitive programming repetition to sport, saying that if you stop training for months the skill goes rusty. Quote: "Stop training for months, and the skill goes rusty." (https://yespress.io/walden-yan)
- Walden Yan eventually became a co-founder and chief product officer of Cognition, the company behind Devin. (https://yespress.io/walden-yan)
- Cognition was founded in 2023. (https://yespress.io/walden-yan)
- With William S. Moses, Walden Yan co-authored a 2020 paper on neural program synthesis. (https://yespress.io/walden-yan)
- After the contest, Walden Yan said he wanted to explore machine learning and website design, then put his programming ability to good use. (https://yespress.io/walden-yan)
- Walden Yan described an early Devin moment in which the team was trying to start a MongoDB server, the documentation was difficult, and the humans were stuck. (https://yespress.io/walden-yan)
- Cognition's founding team included Scott Wu and Steven Hao, and they built Devin, which launched in 2024. (https://yespress.io/walden-yan)
- In June 2025, Walden Yan published an essay titled "Don't Build Multi-Agents." (https://yespress.io/walden-yan)
- Walden Yan wrote that actions carry implicit decisions, and conflicting decisions carry bad results. Quote: "Actions carry implicit decisions, and conflicting decisions carry bad results." (https://yespress.io/walden-yan)
- Walden Yan called the larger discipline of deciding what information an agent needs at each step and how to preserve it over a long run "context engineering." (https://yespress.io/walden-yan)
- Walden Yan's 2025 essay suggested a single agent working in sequence with its context kept continuous as a straightforward starting point, and noted that Cognition had experimented with smaller models to compress history when it grows too large. (https://yespress.io/walden-yan)
- Ten months after "Don't Build Multi-Agents," Walden Yan published "Multi-Agents: What's Actually Working." (https://yespress.io/walden-yan)
- Walden Yan described a pattern pairing Devin with a dedicated reviewer that starts with a clean view of the diff, free from the coding agent's long trail of attempted fixes. (https://yespress.io/walden-yan)
- Walden Yan described a pattern giving a working model a "smart friend" it can consult on difficult decisions. (https://yespress.io/walden-yan)
- Walden Yan's follow-up essay reported a failure in which a smaller primary model in one experiment did not reliably know when to seek the stronger model's help. (https://yespress.io/walden-yan)
- As chief product officer, Walden Yan described in an interview with Anthropic an ambition for an agent that could take higher-level goals, propose its own tasks and coordinate an engineering team to execute them. (https://yespress.io/walden-yan)
- In July 2026, Walden Yan announced that the TierZero team was joining Cognition, pointing to the work of keeping software running after it ships. (https://yespress.io/walden-yan)
- Walden Yan took questions about Devin at a 2025 Devin Meetup Tokyo event for people putting coding agents to work. (https://yespress.io/walden-yan)
- Walden Yan's LinkedIn profile is at https://www.linkedin.com/in/waldenyan. (https://yespress.io/walden-yan, https://www.linkedin.com/in/waldenyan)
- Walden Yan's X profile is at https://x.com/walden_yan. (https://yespress.io/walden-yan, https://x.com/walden_yan)
- Walden Yan's GitHub profile is at https://github.com/walnutwaldo. (https://yespress.io/walden-yan, https://github.com/walnutwaldo)
- Walden Yan is a co-founder of Cognition, the company behind Devin, an autonomous software-engineering agent. (https://ai.engineer/speakers/walden-yan)
- Walden Yan taught himself programming through YouTube videos. (https://ai.engineer/speakers/walden-yan)
- As a first-year Harvard student in 2020, Walden Yan earned a gold medal at the International Olympiad in Informatics, finishing 19th among nearly 400 competitors. (https://ai.engineer/speakers/walden-yan)
- Walden Yan estimated that he had spent more than a thousand hours practicing competitive programming. (https://ai.engineer/speakers/walden-yan)
- After his IOI gold medal, Walden Yan planned to step back from contests to explore his studies, including machine learning and website design. (https://ai.engineer/speakers/walden-yan)
- Through MIT PRIMES, Walden Yan collaborated with William S. Moses on token pairing for neural program synthesis. (https://ai.engineer/speakers/walden-yan)
- Walden Yan and William S. Moses 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. (https://ai.engineer/speakers/walden-yan)
- The TokPair-Model project repository reports 99.24% evaluation accuracy, compared with a previous state-of-the-art result of 95.80%, while using fewer parameters. (https://ai.engineer/speakers/walden-yan)
- Walden Yan and his collaborators created an interpreter that corrected bugs in AlgoLisp's built-in executor. (https://ai.engineer/speakers/walden-yan)
- Walden Yan's repository contains the implementation, dataset preparation, and training and evaluation workflow for the token pairing research. (https://ai.engineer/speakers/walden-yan)
- At Harvard, Walden Yan studied computer science and economics and explored cryptographic applications. (https://ai.engineer/speakers/walden-yan)
- In 2022, Walden Yan 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. (https://ai.engineer/speakers/walden-yan)
- Walden Yan's public projects include EnhanceAI's React wrapper, which integrates AI autocomplete into web application text inputs. (https://ai.engineer/speakers/walden-yan)
- Walden Yan recounted an early Devin experiment in which the team struggled to start a MongoDB server and he asked their rough prototype to get it working. (https://ai.engineer/speakers/walden-yan)
- In the early Devin experiment, the prototype ran terminal commands and deleted files while the team watched with some concern, and eventually the server worked. (https://ai.engineer/speakers/walden-yan)
- Cognition introduced Devin in March 2024, equipping it with a shell, editor, and browser in a sandboxed development environment. (https://ai.engineer/speakers/walden-yan)
- In his June 2025 essay on context engineering, Walden Yan argues that an agent needs the history of relevant conversations, tool results, and decisions. (https://ai.engineer/speakers/walden-yan)
- Walden Yan recommended a single continuing agent as the practical default for agent design. (https://ai.engineer/speakers/walden-yan)
- For longer tasks, Walden Yan proposed compressing an agent's history into consequential details, events, and decisions, while acknowledging that identifying what to preserve requires substantial engineering. (https://ai.engineer/speakers/walden-yan)
- By April 2026, Walden Yan had revised his recommendation around patterns Cognition had found useful, in which additional agents could contribute intelligence while writes stayed single-threaded. (https://ai.engineer/speakers/walden-yan)
- Walden Yan described a fresh-context review pattern in which a reviewer examines the implementation without inheriting the coding agent's accumulated assumptions and long working history. (https://ai.engineer/speakers/walden-yan)
- 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. (https://ai.engineer/speakers/walden-yan)
- Walden Yan describes better results when both models were capable frontier models, with consultation directed toward their different strengths in debugging, visual reasoning, or testing. (https://ai.engineer/speakers/walden-yan)
- In welcoming TierZero to Cognition in July 2026, Walden Yan credited Anhang, Yun, and their team's automation work and described plans to bring it into Devin. (https://ai.engineer/speakers/walden-yan)
- Walden Yan emphasized catching incidents before they spread and maintaining system health over time, with the aim of reducing the time engineers spend handling incidents. (https://ai.engineer/speakers/walden-yan)
- Walden Yan competed for the United States of America at the 2020 International Olympiad in Informatics. (https://stats.ioinformatics.org/people/7322)
- Walden Yan won a Gold medal at the 2020 International Olympiad in Informatics. (https://stats.ioinformatics.org/people/7322)
- Walden Yan scored 505.38 points at the 2020 International Olympiad in Informatics, which was 84.23% of the maximum. (https://stats.ioinformatics.org/people/7322)
- Walden Yan ranked 19th out of 343 contestants at the 2020 International Olympiad in Informatics, a relative rank of 94.75%. (https://stats.ioinformatics.org/people/7322)
- Walden Yan scored 100 points on the 2020 IOI task supertrees. (https://stats.ioinformatics.org/people/7322)
- Walden Yan scored 100 points on the 2020 IOI task tickets. (https://stats.ioinformatics.org/people/7322)
- Walden Yan scored 100 points on the 2020 IOI task biscuits. (https://stats.ioinformatics.org/people/7322)
- Walden Yan scored 100.00 points on the 2020 IOI task stations. (https://stats.ioinformatics.org/people/7322)
- Walden Yan scored 73.38 points on the 2020 IOI task mushrooms. (https://stats.ioinformatics.org/people/7322)
- Walden Yan scored 32 points on the 2020 IOI task plants. (https://stats.ioinformatics.org/people/7322)
- Walden Yan is listed on the IOI statistics site as a Contestant. (https://stats.ioinformatics.org/people/7322)
- Walden Yan is listed among the competitive programmers involved with Cognition, the startup behind Devin, the first AI software engineer. (https://codeforces.com/blog/entry/127049)
- Walden Yan's Codeforces handle is walnutwaldo20. (https://codeforces.com/blog/entry/127049)
- Walden Yan's Codeforces title is Master. (https://codeforces.com/blog/entry/127049)
- Walden Yan has IOI 1 Gold. (https://codeforces.com/blog/entry/127049)
- Walden Yan is listed alongside Neal Wu, Scott Wu, Andrew He, Steven Hao, and Adhyyan Sekhsaria as the competitive programmers involved with Cognition. (https://codeforces.com/blog/entry/127049)
- Previous roles include Early Engineer at Cursor; Co-Founder & CEO at DeepReason; Managing Partner at Inverted Agency; MIT PRIMES Researcher at Massachusetts Institute of Technology. (https://www.linkedin.com/in/waldenyan)
- Location is San Francisco, California, United States. (https://www.linkedin.com/in/waldenyan)

## Experience
- Co-Founder & CPO at Cognition - San Francisco Bay Area (Nov 2023 - Present)
- Early Engineer at Cursor - San Francisco Bay Area (Jun 2023 - Aug 2023): Building Cursor (www.cursor.so)
- Co-Founder & CEO at DeepReason (Feb 2022 - Jun 2023): Next-generation security through automated program verification.
- Managing Partner at Inverted Agency (Jan 2020 - Jan 2021): Sourced leads, built web applications and websites, managed digital media presence for clients, managed a team of 14 consultants and engineers. Got acquired
- MIT PRIMES Researcher at Massachusetts Institute of Technology - Cambridge, Massachusetts, United States (Jan 2018 - Jan 2020): Conducted research in cryptography and machine learning. Wrote papers and presented them to peers.
Designed and benchmarked NPS models, beating state-of-the-art. Collaborated on a compiler library for
optimized fully homomorphically encrypted programs integrated with LLVM.

## Education
- Harvard University - Bachelor's degree, Computer Science and Economics - (Jan 2020 - Jan 2024)

## Sources
- https://www.linkedin.com/in/waldenyan
- https://seas.harvard.edu/news/2020/10/coding-gold
- https://cphof.org/profile/codeforces:walnutwaldo20
- https://codeforces.com/profile/walnutwaldo20
- https://cognition.com/blog/dont-build-multi-agents
- https://en.wikipedia.org/wiki/Cognition_AI
- https://latent.space/wiki/walden-yan-7c572c23/
- https://yespress.io/walden-yan
- https://ai.engineer/speakers/walden-yan
- https://stats.ioinformatics.org/people/7322
- https://codeforces.com/blog/entry/127049
- https://latent.space/p/cognition
- https://github.com/walnutwaldo
- https://x.com/walden_yan
- https://www.linkedin.com/in/waldenyan?miniProfileUrn=urn%3Ali%3Afsd_profile%3AACoAAC1PHewBBkvpJyr7GRyFYLzgX5vJA7VZB00
- https://www.linkedin.com/company/artificial-analysis/posts
- https://www.linkedin.com/in/alexandra-mysak-00883445?miniProfileUrn=urn%3Ali%3Afsd_profile%3AACoAAAmOPggBvQQot1yc_iT6GUo1-WPsbw8Pyzc
- https://www.linkedin.com/company/cognition-ai-labs/posts
- https://www.thecrimson.com/article/2024/3/27/thiel-fellowship-startups-2024/
- https://github.com/openai/swarm
- https://github.com/microsoft/autogen
- https://cognition.ai/blog/dont-build-multi-agents…


Researched with Autumn (https://www.autumn.ai) from public sources. Updated 2026-11-04. Profile: https://www.autumn.ai/resources/cognition/people/walden-yan · Everyone at Cognition: https://www.autumn.ai/resources/cognition/people