Comparative Cognition

Most AI thus far has been deployed with an oddly isolationist architecture: one employee, one terminal, and more often than not, one model. Intelligence is consumed individually, augmenting one person at a time. But organizations have never utilized human intelligence this way. They combine people with different abilities, give them shared context, divide work among them, and coordinate the result. The company, not the individual employee, is the unit that produces.

For two years, we’ve believed machine intelligence will follow a similar path: from local to cloud, from individual to collaborative, and from assistance to execution. Many coding products inherited the assumptions of the personal-computing era, built around an IDE or CLI, one person’s context, and often one model provider. 

Cognition took a different approach. Devin was designed as a cloud agent before the term existed: persistent rather than session-bound, able to operate across systems and coordinate work across agents.

This was easy to miss when Devin launched. The obvious story was an “AI software engineer,” but Cognition had built planning, reasoning, tools, codebase search, and multi-agent infrastructure around the model. As we wrote in May 2024, the application was “thick”: its job was not simply to expose intelligence, but to turn intelligence into work.

When we first invested in Cognition in 2024, six financing rounds ago, we saw the company as the earliest and clearest expression of two beliefs we held about where machine intelligence was heading. First, the frontier would remain jagged, creating gains from trade among models with different abilities. Second, collaborative cloud intelligence would matter more to companies than individual local intelligence.

The Frontier is Jagged

The frontier is advancing quickly, but unevenly. At any given moment, one model may lead on long-horizon agentic work, another on cost per unit of quality, and others on grounding, tool use, or particular forms of reasoning. Those advantages shift constantly. The model best suited to plan a migration across two million lines of code may be an expensive choice for the ten thousand edits that follow, and the wrong choice for triaging an incident at 3:00 am.

If frontier models were interchangeable, enterprises could simply choose a provider and standardize around it. They are not. David Ricardo’s insight was that the gains from trade come from difference: even a country that is better at producing everything benefits from specialization and exchange. Model labs likewise have different comparative advantages, and those differences often matter more for an actual task than on a general benchmark. As the frontier moves, so do the advantages.

This creates an advantage no single-model provider can fully capture. The best harness can use the right model for each subtask, preserve context across the whole job, and change the mix as the frontier moves. A provider-native agent can improve quickly, but will naturally favor its own intelligence. Cognition can draw from the frontier as a whole while also developing intelligence of its own where it has comparative advantage. SWE-2.0 is one example, pushing the frontier on efficiency and intelligence per watt.

The Theory of the Firm

Enterprises ultimately care about finished work, not intelligence in isolation. Turning intelligence into work requires a substantial application layer: Devin must assemble context, plan and delegate work, use tools, verify outputs and know when to involve a human. Increasingly, hard tasks look less like one agent reasoning continuously and more like a firm, with work divided across different forms of intelligence.

This does not mean simply automating humans out of the loop. As we wrote in 2024, Devin was an early example of man-machine symbiosis, with agents executing and coordinating work while humans set direction, supervise and apply judgment. This gives each person leverage over far more work than they could accomplish alone. Human direction and judgment become inputs into the broader system, and the advantage comes from organizing human and machine intelligence together.

Coase asked why firms exist even when markets can allocate resources through exchange. His answer centered on the cost of coordinating complex work, and machine intelligence faces a similar problem. As more intelligence becomes available, the difficulty increasingly lies in coordinating it effectively. A large organization does not need ten thousand isolated copilots, each understanding one employee’s terminal session. It needs agents that can share context and work together across the systems where the company operates.

Context is only part of the challenge. Codebases, documentation, permissions and prior decisions encode how a company actually works; harness engineering determines how agents use that knowledge reliably. Cognition’s Devin Security Swarm, which coordinates specialized agents around a single job, is one example. The work itself already spans people, systems and time; the intelligence executing it increasingly needs to operate across those boundaries as well. In that sense, the useful unit of intelligence becomes the company rather than the individual.

We have seen a version of this before at Palantir, where much of the hard work was in making powerful underlying technology actually useful inside complex organizations. Software engineering was a natural starting point for Cognition because the work was digital, the environment could be instrumented and the output could be tested. But the architecture extends well beyond software.

Cognition × 8VC

When Devin was introduced to the world, there was little to underwrite conventionally. The product could complete some tasks impressively and stumble on others, while the underlying models were not yet capable of much of the work Cognition was designed to do.

We had known Scott for nearly a decade, beginning at Addepar, and had a close view of the team he was assembling. Scott was a competitive-programming prodigy, beating competitors twice his age as a kid, and remains, in his own words, “salty”: intensely competitive, hates to lose, and goes all-in when the problem is worthy of the effort. That proximity mattered enormously. In the early days, the team lived and worked together out of a house in Atherton. The founders had deep respect for exceptional engineering talent and set an almost absurdly high bar from the outset, interviewing thousands of candidates and making only a handful of offers.

It now seems like every week another engineer with their own lore joins Cognition. The culture is intense, curious and urgent, but not self-serious: the same team racing to build the future of software engineering is happy to run a campaign declaring that “Devin is finally good,” poking fun at how far the product had to come.

At the turn of 2025, we led the Series B at $4 billion. By then, the counterargument had become stronger: the labs were shipping agents of their own, models were improving quickly, and it was reasonable to ask whether an independent layer between the enterprise and the model provider would be squeezed out. We believed the opposite. Better models expanded what agents could do, making longer and more valuable work possible. A jagged frontier increased the value of choosing among them, while enterprise adoption made context, verification and trust more important. The layer above the models was becoming more valuable, not less.

By May of this year, enterprise usage had grown more than tenfold since January, and we co-led the next round at $26 billion. The numbers were evidence of something we had believed before there were numbers: enterprises would buy finished work, not merely intelligence, and a company able to draw from the frontier as a whole would have an advantage over one tied to a single part of it.

We were not perfectly right in 2024. Cloud agents took longer to become useful than the first demonstrations suggested, and the boundary between agent companies, model providers and incumbent software remains unsettled. But the core observations have held: AI applications require a thick layer around the model; different intelligences are better suited to different work; and organizations need coordination more than they need another isolated assistant.

Software engineering was the natural starting point for Cognition. The ambition is much larger: to change how companies organize human and machine intelligence.

‍