Cognition

What Cognition is

Cognition is the company that introduced Devin on 12 March 2024, describing it as the first autonomous software engineer^[1]. Rather than selling a coding assistant that suggests lines inside an editor, the company positions Devin as an agent that takes on a task and carries it through, and it frames that work as something done together with people rather than in place of them.

How the company frames its work

The company states that it believes the purpose of technology is to expand human capacity, not by replacing meaningful work but by working alongside people as an exponential collaborator^[1]. It says it builds tools for software creation because software shapes nearly every aspect of modern life^[1]. Its stated aim is to help engineers operate more like architects: strategizing, designing systems and focusing on problem solving while agents handle repetitive engineering work^[1]. That framing is worth keeping in mind when reading claims about what the agent can do, since the product and the marketing come from the same place.

Devin

According to the company, Devin plans, writes, tests and ships production code on its own, working inside a customer's codebase and the tools the team already uses^[1]. The emphasis on the existing codebase and existing tools matters: the pitch is not that a team moves to a new environment, but that the agent is dropped into the one the team already has. Whether an agent's output is genuinely usable is a separate question, and it is the question Cognition's own benchmark tries to answer.

FrontierCode

Cognition also publishes FrontierCode, a benchmark that asks whether a model writes code a maintainer would actually merge, rather than only whether it passes unit tests^[1]. The distinction is practical. A model can satisfy a unit test and still produce a patch a human reviewer would reject — awkward structure, poor fit with the surrounding code, or a fix that solves the stated case and breaks the neighbouring one. FrontierCode 2.0, published on 27 October 2026, adds more diverse and harder tasks from more repositories, graded by open-source maintainers on whether agent output is actually mergeable^[2]. Grading by maintainers rather than by an automated pass/fail is the notable part: the judgement being measured is the one a reviewer makes before accepting a change.

Growth beyond the coding agent

On 22 October 2026 the company announced it had opened an office at The Gilbert in Finsbury Square, London, after its UK team grew by more than 200% in six months^[2]. On 28 October 2026 it announced that Corridor is joining Cognition to take Devin Security further^[2]. On 4 November 2026 Cognition introduced SWE-2 Flash, described as a fully American coding model aimed at low cost and high speed^[2]. Together these announcements sketch a company widening in three directions at once: a physical presence outside the United States, a security-oriented line of work under the Devin name, and its own model family sitting underneath the agent rather than relying only on models from elsewhere.

Where it sits

Cognition belongs to the group of companies that, from 2024 onward, argued that software work could be delegated at the level of a task rather than a keystroke. Its public output covers three things at once: the agent itself, the benchmark used to argue about agent quality, and the underlying models. Publishing a benchmark that measures mergeability by maintainer judgement is also a way of setting the terms of the argument — if the metric is whether a human would accept the change, then the agent is being judged on the same standard as a colleague's pull request^[1][2]. How far that standard has actually been met is something the company's own published results speak to; the claims above are the company's description of its product and its measurements, not an independent assessment.