Cognition's $26B Round: Devin Writes 89%. Source the 369-Person 11%.

Cognition just closed more than $1B at a $26B valuation, co-led by Lux Capital, General Catalyst, and 8VC. The number that should actually reset your sourcing plan is buried in the company's own Series D post: Devin now writes 89% of the code committed at Cognition, up from 13% in December 2025. The headcount sits at 369. If you are a recruiter or eng leader looking at this round and thinking "they must be hiring like Anthropic," you are sourcing the wrong company.

The math: $26B over 369 people

Tracxn pegged Cognition at 369 employees as of April 30, 2026. The company hit a $492M annualized run rate in May 2026, up from $37M twelve months earlier. That is a 13x revenue jump on a headcount that, by every public account, barely moved.

The founders are telling you why explicitly. Scott Wu, Steven Hao, and Walden Yan have repeated the phrase "small and talent-dense" across every post-raise interview. The Series D blog post names the founding alumni nodes: Cursor, Scale AI, Lunchclub, Modal, Google DeepMind, Waymo, Nuro. The March 2024 ten-person team held ten IOI gold medals between them, including Gennady Korotkevich and Andrew He. That is not a hiring funnel. It is a roster.

89%

So when "Cognition AI hiring" trends in your sourcing channels this week, the first thing to internalize is that the company is not going to absorb 800 engineers off the open market in the next 18 months. The structural answer is closer to 80. Maybe 120.

What the 11% actually does

The remaining 11% is not generalist backfill. It is the orchestration layer that makes the other 89% tradeable to Citi, Mercedes-Benz, Goldman Sachs, Elevance, Dell, Santander, and US government buyers.

Cognition does not train a frontier model. The $26B valuation is a bet on harness design, agent orchestration, evals, and the workflow that wraps somebody else's weights. That collapses the human hiring profile to a handful of specialties:

• Agent reliability engineers. Cognition's March 2026 paper made the case that an agent succeeding on 90% of tasks but failing unpredictably on the other 10% is unacceptable as an autonomous system. Reliability improvements lag accuracy improvements at roughly half the rate. The people who close that gap look like SRE crossed with formal methods crossed with evals.
• RL and post-training engineers. Not pretrainers. People who can shape behavior on top of frontier models with reward design and rejection sampling.
• Harness and tool-use engineers. The plumbing between the LLM, the IDE, the shell, the browser, and the codebase. Most of this skill set was invented at Anysphere, Cognition, Magic, and the Claude Code team in the last 24 months.
• Eval leads. SWE-bench is the obvious one. The interesting hires are people who built private internal evals at OpenAI, Anthropic, or DeepMind and can do it again from scratch.
• Infra and security. Devin runs code on customer infrastructure. Itaú resolves 70% of its security vulnerabilities automatically with Devin. Mercedes-Benz compressed an eight-month modernization project to eight days. That sales motion only works if the security story is bulletproof, and bulletproof requires real humans.

Notice what is not on that list. There is no "senior backend engineer." No "staff frontend." No generalist SWE III. Scott Wu has publicly framed Devin's current level as "between a junior engineer and a mid-level engineer." That is exactly the band Cognition has stopped hiring.

The sourceable pool is a list, not a funnel

Here is the part that breaks most ATS workflows. The talent Cognition will actually consider is small enough to enumerate by hand. A few hundred IOI-tier competitive programmers over the relevant decade. A few hundred more agent-harness practitioners who have shipped something real. The overlap is brutal.

For competitive programming signal, the sourceable surfaces are Codeforces top-rated lists, ICPC World Finals rosters from roughly 2014 through 2022, IOI medalist rosters, and Putnam fellows. Walden Yan was a Thiel Fellow before Cognition, so Thiel class lists are in scope too. Most recruiters do not have these as searchable fields. LinkedIn certainly does not.

For agent-harness signal, the named pools are concrete: Anysphere/Cursor, Anthropic's Claude Code team, Magic.dev, Imbue, Poolside, Codeium and Windsurf alumni, Modal, the OpenAI Codex team, Replit Agent, Sourcegraph Cody, All Hands AI (OpenHands), and Reflection AI. The SWE-bench leaderboard author list is a goldmine. So is the LangChain agent-evals contributor graph.

This is exactly the search that boolean strings fail at. "Member of Technical Staff" AND "LLM" AND "agents" on LinkedIn returns either everyone who has tweeted about Devin or nobody who has actually built one. Plain-English search across GitHub, LinkedIn, and the open web is the only thing that resolves it, which is why we built Refolk: you describe the person you want ("MTS-level engineer who has shipped an agent harness, ex-Cursor or ex-Anthropic, currently in SF or NYC") and get a ranked shortlist instead of 4,000 false positives.

When we ran a representative pass on the Refolk index for US senior+ engineers tagged with LLM and Agents skills, the top current employers concentrated tightly: Anthropic (4), Microsoft AI (2), Magic (2), and one each at OpenAI, Imbue, Contextual AI, Resolve AI, Reve Image, and Salesforce. Almost all SF Bay Area or NYC. That is the real shortlist of MTS-types who could plausibly walk into Cognition. It is also, conveniently, the shortlist Cognition's competitors should be defending right now.

Windsurf is the underrated alumni node

Cognition acquired the remaining assets of Windsurf in mid-2025 after the Google deal carved out its leadership. That absorbed an entire IDE team plus the pre-acquisition Exafunction crew. Anyone with "Windsurf" or "Exafunction" on their profile right now is either inside Cognition or recently out the door.

The recent-departures cohort is the highest-signal sourceable group outside the founding circle. They have seen Devin's internals. They speak the harness vocabulary. They are not bound by the Cognition retention package that the Series D presumably refreshed. They are also rare enough that LinkedIn boolean does not find them reliably, because half list "Codeium" and half list "Windsurf" and a few list both with inconsistent date ranges.

The Cognition profile is not a funnel. It is a list of about 300 names, and most of them are not on LinkedIn the way you think they are.
</pull>

## The comp story changed the day Devin hit 89%

Here is the second-order effect nobody is pricing in. When a coding company's coding agent writes 89% of the coding company's code, the marginal human hire has outsized leverage on a $492M run rate. That math justifies total compensation well above the Anthropic and OpenAI MTS bands, which themselves cleared $1M cash and equity for top profiles in 2025.

If you are an AI coding startup recruiter trying to poach out of Cognition, assume the counter is brutal. If you are recruiting *into* Cognition from Anthropic or Magic, assume you are competing on something other than cash, because cash is unlikely to be the gap. The interesting levers are scope, autonomy over the harness, and proximity to the founders.

If you are a Series B or C AI coding startup hiring against this round, the realistic play is the Windsurf alumni cohort and the Cursor mid-tier, not the Cognition core. The Cognition core is not interview-available.

## The Scott Wu contradiction worth quoting

Wu has spent the post-raise press cycle arguing that AI agents should not replace humans and that humans must retain final decision-making authority. He says this while Devin writes 89% of his company's code and he runs a 369-person org generating $492M ARR.

That contradiction is the most useful framing device for any recruiter pitching candidates on Cognition. The company is not selling "we replaced our engineers." It is selling "our engineers are 9x leveraged." That pitch lands very differently with a staff engineer who is tired of being one of 400 people on a platform team at Stripe. It is also a much easier pitch to make if you can show the candidate exactly who else is on the team and what they shipped last quarter, which is the kind of map [Refolk](/) builds in a few minutes instead of a few weeks.

## What to do this week

If you run sourcing for a competitor (Magic, Poolside, Reflection, Anysphere, Sourcegraph, Replit, All Hands), the move is to enumerate the recent Windsurf and Exafunction departures and the SWE-bench author list before Cognition's expanded recruiting budget gets pointed at them. The Series D closed in late May. The retention conversations inside Cognition will land in June and July. The window where ex-Windsurf engineers are most receptive to a competing offer is roughly the next 60 days.

If you are an enterprise eng leader thinking about a Devin pilot, the recruiting implication is different but real. Itaú and Mercedes-Benz did not buy Devin in a vacuum. They restructured around it. That restructure requires a handful of internal agent-reliability and evals hires you do not currently have. The autonomous coding agent talent market is the same one Cognition is fishing in, and you will lose the bid on the top of the pool. Source one or two layers down: people who have contributed to OpenHands, who maintain agent-evals repos, who are second authors on agent papers from the last 18 months.

If you are a founder thinking the Cognition story validates a 12-person engineering org at Series A, read the founding team list one more time. Ten IOI gold medals. Korotkevich. Andrew He. Yan as a Thiel Fellow. That is the prerequisite, not the consequence. Refolk is useful precisely because most teams cannot start with that roster and need to assemble the equivalent from a much wider pool of Cognition Series D engineers, ex-Cursor staff, and adjacent agent practitioners without burning a quarter on boolean searches.

The Cognition round is not a hiring story. It is a sourcing story disguised as a hiring story. The companies that win the next year are the ones that figured out the difference this week.

## FAQ

### How many people will Cognition actually hire after the Series D?

Cognition has not announced a target, but the structural evidence points to small numbers. Headcount sat at 369 in April 2026 against a $492M run rate, and Devin writes 89% of internal code. A doubling of headcount over 18 months would be aggressive by Cognition's own standards, and even that only adds a few hundred roles, almost all in agent reliability, evals, RL, infra security, and enterprise forward-deployed engineering. Generalist SWE hiring is unlikely to be a major line.

### Where do I actually find Devin-adjacent engineers if LinkedIn does not surface them?

The high-signal surfaces are SWE-bench leaderboard authors, the LangChain and agent-evals open-source contributor graphs, Codeforces and ICPC rosters for the competitive-programming profile, and the recent-departures lists from Windsurf, Cursor, Magic, Imbue, and the Claude Code team. Most of these names do not have searchable LinkedIn titles that match what you want. Plain-English search across GitHub, LinkedIn, and the open web (the problem Refolk solves) is the cleanest way to compile the list without a month of manual work.

### Is Cognition's $26B valuation defensible if Devin keeps improving?

That is the interesting tension. Devin writing 89% of Cognition's code compounds the company's own velocity, but it also implies the broader market for autonomous coding agents will commoditize whatever Devin does today. The defensibility argument inside Cognition rests on harness quality, enterprise distribution (Citi, Goldman, Mercedes-Benz, Santander, Elevance, US government), and the reliability gap the March 2026 paper described. None of those are model-weight moats. All of them are talent moats, which is why the hiring profile matters.

### Should I tell candidates Devin will replace them?

No, and Scott Wu's own framing is the script. Devin sits between a junior and mid-level engineer in capability. Cognition's pitch to its own hires is leverage, not replacement: a 369-person company running $492M ARR means each engineer's output is multiplied by the agent layer underneath them. Candidates who hated being one of 400 platform engineers at a megacap respond to that pitch. Candidates who want to manage people do not, and that is fine. Sort on it early.