Cognition AI’s Devin-kun deal: Tokyo now modernizes legacy code with agentic software

Japan has a new favorite kind of coworker: an AI software engineer so autonomous it can write, debug, and deploy code inside the tools a team already uses. Cognition AI, the San Francisco startup behind Devin, is making that bet practical by opening its first Asia office in Tokyo in April, then planning to make Singapore its Asia-Pacific headquarters later this year. The pitch is straightforward, and the urgency is not: Japan is confronting aging digital infrastructure built on legacy code while its workforce shrinks.

Cognition President Russell Kaplan put the framing bluntly in early June: “Japan was our first or second most popular country in terms of user engagement overall.” He also tied the demand to hard constraints, saying, “The needs are real, especially in critical infrastructure and government,” and that “The country is running on aging infrastructure with a declining workforce.” Japan’s demographic reality is doing the heavy lifting here. The country has the world’s oldest population, with almost 30% of its residents over the age of 65, and its working-age population is projected to decline by over 30% between now and 2060. That demographic squeeze maps directly onto a talent shortage. In 2023, Japan’s Ministry of Economy, Trade, and Industry (METI) estimated Japan would face a shortage of 789,000 software engineers by 2030.

Put those together and you get a market that rewards speed. Cognition is effectively selling time. It pointed to Sapporo’s city government, which needed to modernize over one million lines of legacy code to meet a national IT compliance mandate. Kaplan estimated that the work would have normally taken 200 engineering months. Using Devin, Sapporo’s engineers completed it in roughly a quarter of that time. That’s the kind of efficiency claim that boards and city CIOs can’t ignore, because it translates “AI coding” into something less abstract: a modernization timeline and a compliance deadline.

Japan’s “AI agents” moment is also getting cultural traction in a surprisingly specific way. Even before Cognition officially launched in Japan, Devin was already going “viral” there. Kaplan said there was a debate about the correct honorific for Devin, referring to the suffixes attached to names to designate social hierarchy. What the community settled on was “Devin-kun.” It’s a small detail, but it signals something bigger for a company trying to sell software into conservative environments: users are talking, joking, and testing. They are also treating the agent as a peer, not a novelty.

The timing matters because Japan is also a favored beachhead for U.S. AI companies looking to expand globally. OpenAI and Anthropic opened their first international offices in Tokyo. Microsoft, Alphabet, and other hyperscalers have committed billions to Japanese data centers. Japan also secured access to Anthropic’s Mythos model via Project Glasswing, which was designed to help key companies and critical institutions fix security vulnerabilities, and three of Japan’s largest banks - MUFG, Mizuho, and Sumitomo Mitsui - were among those granted entry. That access was quickly shut off after the U.S. barred all foreigners from using the model in mid-June.

This is where Cognition’s strategy gets sharper. Kaplan suggested Japan is comfortable sticking with U.S. AI “due to the country’s investment and close relationships with American AI labs.” He also said Japan has “disproportionately invested in working closely with U.S. companies to influence the roadmaps of those companies to meet local domestic needs.” The subtext for executives is that “where you deploy” and “who you partner with” can affect not just performance, but priorities. Japan also has multilingual barriers, and Kaplan noted that low English proficiency “has led to a bit more isolation for some businesses in Japan.” Devin can work in Japanese while collaborating through the agent with teams on the other side of the world, which matters when the bottleneck is communication, not compute.

Cognition’s technology is positioned as a full teammate, not a chatbot. Give Devin a task, and it codes, debugs, and deploys code autonomously inside the tools an engineering team already uses. The company describes Devin as an early example of “AI employees” or agents fully integrated into workplace tools like Slack, where employees can assign tasks without constant prompting. The business side of this is scaling fast. Cognition AI, founded in 2023, raised more than $1 billion in a new funding round in late May, valuing the startup at $26 billion, more than doubling its valuation from a September 2025 round. At the time of the raise, Cognition’s annualized run rate reached $492 million, up from $37 million a year earlier.

Now comes the second-order impact, and it hits a nerve in global software services. Cognition’s coding tools can look, to some investors, like an existential threat to existing programmers and software engineers, particularly in places such as India that have traditionally dominated back-office work. The prospect of AI agents performing that work at a fraction of the cost has rattled investors. Shares in Infosys, Wipro, Tata Consultancy Services, and HCLTech have all fallen between 30% and 40% over the past 12 months. Kaplan said he isn’t worried about India’s ability to adapt to AI. He argued that the job can become more “fun and impactful” because engineers get someone working on a specific part of a project while also opening opportunities for promotion as they manage a whole team of AI agents. He also said the companies Cognition works with are using productivity gains to become more ambitious.

Zoom out and you see why Japan is not just a “nice early customer” story. Cognition is also finding another advantage: compute. Kaplan said compute is finite and that demand at Cognition is doubling roughly every seven weeks. Geographically diverse teams allow compute to be used during off-peak hours on Wall Street and Silicon Valley. He framed it simply: when people are at work in Japan, people in New York are asleep, creating efficiency for an AI company. For executive teams elsewhere, that’s a reminder that agentic software is not only a labor story. It is also a capacity planning, regulatory, and infrastructure story.

Cognition is still building its Asia footprint, with Malaysia as another unexpected growth market. Kuala Lumpur has become a regional software engineering hub, driven by a large English-speaking talent pool, lower operating costs, and proximity to the rest of Southeast Asia. Kaplan described the engineers his team encountered there as among the most skilled in the world at managing AI agents. Cognition has launched an Applied AI Engineering program in Malaysia to identify top engineers who excel at directing agents, training them to teach entire teams how to work effectively with AI. Kaplan also said Cognition is looking closely at South Korea and Australia as possible Asia-Pacific expansion markets.

For founders, investors, and enterprise leaders, the stakes are clear: if Japan can cut modernization cycles by roughly a quarter while tackling a projected shortage of 789,000 software engineers by 2030, then “AI coding” stops being a pilot category and becomes an operating model. The winners in this phase will be teams that treat AI agents as infrastructure for delivery speed, not just tools for clever demos.