Jane Street

About

Our team is small and talent-dense. The founding team holds 10 IOI gold medals and includes leaders and builders who have worked at the cutting edge of applied AI at companies like Google DeepMind, Scale AI, Cursor, Lunchclub, Modal, Waymo, and Nuro. Of our initial ~35 team members, 21 of them had founded a company.

Wer'e developing a suite of agentic tools - led by Devin, our AI software engineer - that operate across the entire software lifecycle. Our mission is to create agents that can reason, remember, and execute complex work over long horizons, as true teammates to human engineers.

We're still in the early innings - our hardest challenges lie ahead. If you're excited to solve some of the world's biggest problems and build AI systems that can reason and learn, we'd love for you to explore our careers.

Message from Grant

I visited the nascent Cognition team not too long after it was founded, and it immediately left an impression. For context, I knew about them because a college friend of mine, Russell Kaplan, is now president there. Throughout the last decade, more than anyone else I know, Russell has had an uncanny ability to predict the major trends in AI a few years before they become mainstream. He's also one of the most savvy and ambitious people I know, so when he pointed his attention towards this new company, naturally, I was intrigued.

In that first visit, what struck me is how the sheer concentration of both talent and hustle is leagues above what you see in most companies, even start ups. It's something like the software engineering equivalent of a close-knit team of athletes training for the Olympics.

I visited again more recently in the context of this partnership with 3b1b, and though the days of everyone working from one house are now long gone, I was pleased to see that through their growth, the talent concentration remains as dense as ever.

The Hard Problem We're Solving

Modern software development isn't a sequence of isolated keystrokes, but a multi-stage process that spans understanding existing systems, planning changes, writing and reviewing code, testing, documenting, and maintaining software over time. Our goal is to build AI teammates that can operate across this entire lifecycle, not just assist at a single step.

That's why Cognition is built as an agent lab: a place where research and engineering come together to develop a shared "agent brain" that powers a suite of products across the SDLC. This includes Devin, an autonomous software engineer; Windsurf, a next-generation IDE for creative 0-to-1 coding; Devin Review, which brings deep, contextual reasoning into code review; and DeepWiki, which helps teams understand and plan changes in large, complex codebases.

Our approach centers on pushing agent capabilities further along three dimensions:

• Long-horizon planning and reasoning, where agents construct and adapt multi-step strategies before and during execution - not just react to the next prompt.
• Deep, persistent codebase understanding, enabling agents to build a mental model of large systems, recall prior decisions, and reason about tradeoffs across files, services, and time.
• Autonomous execution with feedback loops, where agents write code, run tests, review changes, and iterate within real engineering environments and workflows.

By grounding these capabilities in real developer tools and workflows, we're building agents that don't just generate output - they do work.

We envision a future where engineers collaborate with a fleet of AI teammates that deeply understand their code, their tools, and their goals - freeing humans to focus on architecture, creativity, and the hardest problems, while agents handle execution at scale.

See how engineering teams are working with Devin today.

Featured Work

SWE 1.6 reflects how we think about building AI that actually works in practice - not just on benchmarks, but in the moment-to-moment experience of using it.

SWE 1.6 was post-trained from scratch to eliminate inefficiencies - bloated trajectories, sequential tool calls, and unnecessary terminal reliance. A length penalty in training was key: it discourages overthinking without sacrificing capability, pushing the model toward parallel tool use and faster context retrieval. The result: solve rates hold steady while assistant turns stay flat. Less friction, same intelligence.

That kind of disciplined tradeoff doesn't happen by accident. Internally, this translates into a culture that values:

• systems thinking over surface-level hacks,
• fast iteration paired with technical rigor,
• and building tools we ourselves rely on every day.

SWE 1.6 proves you don't have to choose between speed or craft - and that belief shapes how we build.

Join

What I love about Cognition is the urgency. We're racing to build the future of software engineering, and everyone here feels that weight and excitement. It's intense, but when you see Devin or Windsurf solve problems you didn't think were possible yet, it's the most fulfilling thing in the world

— Theodor Marcu, Head of Product Growth

Cognition is a small, talent-dense team with a deeply competitive, low-ego culture. People move fast, argue passionately about hard problems, obsess over details others ignore, and take pride in making complex systems work at scale. If this sounds like you, we'd love to hear from you!