Devin migrated Cognition's site and recorded its own verification runs

Why it matters

Cognition's central bet depends on agents returning evidence, not just code. Recorded testing could reduce the review burden as engineering teams delegate more production work asynchronously.

Cognition, founded by Scott Wu (@ScottWu46), Steven Hao (@stevenkplus1), and Walden Yan (@walden_yan), used Devin to migrate its marketing site from Astro to Next.js, with the agent testing the result and recording runs that helped reproduce subtle bugs and confirm fixes, according to Jared Palmer (@jaredpalmer) in a thread.

Jared Palmer on X

The timing of the migration itself is unclear, so the episode is best read as an internal case study rather than a new product release. Palmer did not publish the repository, duration, number of files changed, deployment status, or before-and-after performance data. His account also leaves the human contribution unquantified, including who scoped the migration, reviewed the output, and approved the final deployment.

Those omissions limit what the migration proves about autonomy. The testing workflow is still the consequential detail. An agent that changes a framework can produce an impressive diff while quietly breaking navigation, styling, authentication, analytics, or browser-specific behavior. Palmer's account says Devin used recorded executions to return to those subtle failures and verify later fixes, turning the run itself into an artifact an engineer could inspect.

The founders' bet reaches the review stage

Wu and his co-founders built Cognition around delegation rather than inline code completion. Cognition's stated goal is an agent that accepts a software task, works inside the repository and development environment, and returns finished work for review. The founders came out of competitive programming: Cognition says its founding group collectively holds 10 International Olympiad in Informatics gold medals and includes engineers from Scale AI, Cursor, Lunchclub, Google DeepMind, Modal, Waymo, and Nuro.

Wu, a three-time IOI gold medalist who previously co-founded Lunchclub, has described Cognition's aim as making it easier for people to build reliable software while engineers move toward system design and problem definition. The marketing-site migration fits that thesis unusually well. Framework conversions are bounded enough to split into concrete steps, yet broad enough to expose whether an agent can navigate a real codebase, preserve behavior, and catch regressions after the code compiles.

Palmer is also a pointed messenger for the work. His personal site identifies him as Cognition's vice president of engineering. His account comes from Cognition's VP of Engineering, not from one of its founders.

Recording becomes part of the deliverable

Cognition has been adding recorded verification to Devin in 2026. In February, Cognition released Devin 2.2 with desktop computer use, self-verification, and automated fixes. Cognition's computer-use documentation says Devin can install dependencies, start an application, generate a test plan, interact with the interface, and return an annotated video of the run.

Cognition engineer Ido Pesok described the strategic reasoning in a May 29 post. As more agent sessions begin through schedules, automations, and other agents, engineers cannot watch every step live. The output needs enough evidence for someone returning later to judge whether the change is ready for review. Cognition says Devin can also turn recurring test procedures into reusable skills, which matches Palmer's explanation that verification scripts may be stored in a session scratchpad or reused across parallel child sessions.

That workflow addresses a basic problem created by asynchronous coding agents: they increase the amount of code available for review faster than they increase engineers' capacity to understand it. A screen recording does not establish correctness on its own. It can make the review narrower by showing which flows ran, what the agent observed, and whether the reported fix survived the same sequence that exposed the bug.

Cognition's own 2025 performance review draws the boundary plainly. Cognition said Devin worked best on tasks with clear requirements and verifiable outcomes, while ambiguous projects and test logic still required human judgment. A framework migration can sit on either side of that boundary. Acceptance criteria such as passing builds, preserved routes, and stable visual behavior are testable. Decisions about architecture, accessibility, long-term maintainability, and performance tradeoffs remain harder to reduce to a recorded run.

Independent testing has previously shown the cost of getting that boundary wrong. In January 2025, Answer.AI reported 14 failures, three inconclusive results, and three successes across 20 tasks it assigned to an earlier version of Devin. The researchers found that Devin could become stuck on poorly specified or unfamiliar work. That evaluation predates Cognition's 2026 verification releases and does not assess the Astro migration, but it explains why evidence attached to the output has become central to Cognition's product pitch.

Cognition needs proof that scales with the valuation

The internal migration also serves a commercial purpose. Cognition said in May that it raised more than $1 billion at a $26 billion post-money valuation in a round led by Lux Capital, General Catalyst, and 8VC. Cognition also reported $492 million in annualized run-rate revenue and more than tenfold enterprise-usage growth since the start of 2026. Those figures are self-reported, and Cognition has not attached audited financials to the announcement.

At that scale, Cognition is selling an operating model for engineering organizations, not a clever code generator. The Astro-to-Next.js migration gives prospective customers a compact version of that pitch: assign an existing production project, let the agent make the changes, and receive inspectable evidence alongside the code.

Palmer's thread stops short of establishing a zero-touch migration or measuring whether Devin beat an experienced engineer on time, cost, or quality. It does show where Wu, Hao, and Yan are concentrating Cognition's product work. Code generation is already abundant. Cognition is betting that planning, execution, testing, and reviewable proof can make autonomous work dependable enough to merge.