Cognition: Devin Writes 89% of Code. Humans Stay in Charge.

Devin, Cognition’s AI coding agent, now handles 89% of the company’s commits. The startup raised $1 billion at a $26 billion valuation. CEO Scott Wu cuts against the prevailing narrative: these agents aren’t here to replace developers, but to absorb the friction they never wanted anyway.

Key Takeaways

• Devin handles 89% of commits at Cognition, operating between junior and mid-level engineer capability depending on the task
• Cognition raised $1 billion at a $26 billion valuation
• Scott Wu sees AI agents as another abstraction layer in programming history, never as a substitute for human developers

What 89% of Commits Actually Tells Us About Agents Today

The figure came straight from the CEO’s mouth. At Cognition, Devin handles 89% of the company’s code commits. This isn’t a roadmap projection or a controlled experiment result: it’s the daily operational reality at the company that built the agent. There’s a candor in that disclosure that AI companies rarely apply to their own internal tools. For context, see our earlier piece on Horizon: Claude Code Artifacts: Anthropic Opens Sharing.

But the 89% figure deserves unpacking. According to Scott Wu, Devin operates somewhere between a junior and mid-level engineer depending on the task. It excels at the repetitive blocks: maintenance patches, dependency migrations, unit tests, configuration updates. These are exactly the tasks senior developers spend time on without finding any creative satisfaction in.

The remaining 11% is not trivial. It concentrates the architectural decisions, the technical trade-offs, the creativity no agent can replicate yet. Wu is explicit about it: Cognition’s engineers love their craft precisely for those moments. What the agent absorbs is the friction, not the work that matters.

This split points to a model spreading across the industry. As we covered in our analysis of Asana’s acquisition of StackAI, software vendors are repositioning as platforms built for mixed human-agent teams. The question is no longer whether agents will enter enterprise workflows: they already have. The question is where their autonomy stops.

In the short term, companies deploying coding agents will start measuring their own commit ratios. Cognition’s 89% becomes a market benchmark. Teams sitting at 10% will wonder why. Teams reaching 95% will face different questions about what human involvement still means. The full picture is available in Cognition’s annual Devin performance review covering its coding output.

The Abstraction Argument

Scott Wu didn’t need AI to think about programming in terms of abstraction layers. A renowned competitive programmer since childhood, he has spent his career pushing the limits of what code can accomplish. His argument holds together in one statement: every generation of tools has made developers more powerful by absorbing a layer of complexity. AI is simply the latest layer.

He draws on historical precedent. Visual environments abstracted memory management. High-level languages abstracted assembly. Frameworks abstracted network infrastructure. At each step, predictions about the end of the developer emerged. At each step, developers became more numerous and more productive. Wu sees Devin as the direct continuation of that pattern, not a break from it.

The nuance matters: Wu isn’t saying nothing will change. He’s saying what will change is the nature of tasks accessible to any given developer. A solo engineer will be able to run projects that previously required a team of five. A mid-level will be able to deliver senior output on the mechanical parts of a codebase. The skill hierarchy shifts, it doesn’t disappear.

That shift connects to a broader movement. The web is tipping toward majority non-human traffic, a progression we documented in our piece on the rise of agent-generated web traffic. Cognition’s agents don’t just produce code: they generate API calls, database queries, and entire automated interaction sessions with third-party systems.

Over the medium term, this shift in the nature of work will create structural pressure on engineering curricula. Learning to direct an agent, break a problem into delegatable subtasks, and validate automated output is becoming as foundational as writing code itself.

Also on Horizon:

• Asana Buys StackAI for $75M to Build the Agent Workplace
• Non-Human Traffic Is 31% of the Web and Growing Fast
• Vibe: Mistral Turns Le Chat Into a Full Work Agent

Beyond Code: Customer Service, Medicine, and the Control Line

Scott Wu doesn’t limit his vision to software. He anticipates rapid expansion into other domains: customer service first, then medicine. In both cases, agents capable of acting in high-stakes environments, serving end users who may be vulnerable. His position is unambiguous: agents must “always” remain under human direction. That word carries weight.

Customer service is the first target. The reasons are obvious: massive volumes of repetitive requests, relatively higher tolerance for imperfection. But even there, the risk of autonomous drift is real. An agent that misreads a request and acts without escalating to a human creates damage that’s hard to contain. Supervision isn’t a governance detail: it’s an operational safeguard.

Medicine is a different story. Regulatory stakes are incomparably higher, and error tolerance approaches zero in many clinical contexts. Wu knows this. His insistence on permanent human oversight makes full sense here. In a sector where a misinterpretation can cost a life, the question isn’t “can the agent technically do this?” but “who validates every step?”

In the short term, Cognition will deploy in low-risk environments. The $1 billion raise will fund expansion into new verticals, deeper capabilities on complex codebases, and likely specialized offerings for specific industries. The money isn’t for consolidating what exists: it’s for accelerating sectoral reach.

Over the medium term, the real tension will play out in the regulatory arena. Regulators in both the US and Europe are beginning to draw lines around AI system autonomy in critical sectors. Cognition, like every agent AI player, will need to navigate between technological ambition and regulatory compliance. Wu’s position on permanent human control isn’t just an ethical stance: it’s also a strategy for staying ahead of regulation that’s coming.

Follow the story on Horizon.