cognition

a sourced history of cognition, scott wu's company behind the devin software-engineering agent, and its acquisition of windsurf.

cognition is the applied ai lab behind devin — three ioi gold medalists who started in a new york apartment, ran roughly eight pivots, and in march 2024 declared they had built 'the first ai software engineer.' the claim survived its own debunk, a 3-of-20 independent evaluation, and a benchmark it had to disown — and the company still ran the valuation arc $350m to $48b in thirty months, bought windsurf in a weekend, and repositioned itself as the independent agent lab. this is the record of that bet — the medalist founding team, the launch and the debunk, the windsurf twist, and what the agent-era idea actually proved.

• history: the company, sourced and dated — the ioi apartment, the devin launch and debunk, the windsurf weekend, and the $48b arc.
  • the ioi apartment: 2023–early 2024 — wu, hao, and yan, ~eight pivots, a january 2024 incorporation, and the $21m thiel round.
  • the first ai software engineer: 2024 — 'the first ai software engineer,' the 13.86% swe-bench claim, the debunk, $2b, and ga at $500.
  • the reality gap: jan–jun 2025 — answer.ai's 3-of-20, the $500→$20 price collapse, deepwiki, and the ~$4b round anyway.
  • the windsurf twist: jun–aug 2025 — openai's lapsed exclusivity, google's $2.4b talent lift, cognition's monday signature, the 80-hour memo.
  • the independent agent lab: sep 2025 → sep 2026 — $10.2b → $26b → $48b, swe models on cerebras, devin desktop, government contracts.
• people: scott wu, steven hao, walden yan, and the people around cognition — who they are, what they wrote, and what they claim.
  • the founders: wu, hao, and yan — three ioi gold medalists, three very different public profiles, one cap table.
  • the writing: the launch manifestos, the funding memos, the context-engineering essays — cognition's paper trail in its own words.
  • the team: kaplan, the medalist bench, jeff wang, the extended roster, and the backers who underwrote the arc.
  • interviews and talks: cheeky pint, senra, colossus, ai & i — where the claims got said out loud.
  • the philosophy: agents as coworkers, intensity culture, context engineering — the themes the company keeps returning to.
• side quests: the pursuits around the product — lunchclub, the upwork demo, openhands, deepwiki, the swe model family, and the diaspora.
  • the founders' quests: lunchclub, the eight pivots, and the upwork side hustle — the quests the founders brought with them.
  • the company's quests: deepwiki, the oss initiative, the swe models, government work, the guarantee, rsa-260, and the acqui-hire shelf.
  • the ecosystem: openhands, the diaspora, and the community's own record of what devin meant.
• coverage: what the discourse made of cognition — the launch press, the debunk cycle, the windsurf saga, and the funding arc.
  • the launch: the gold-medalist feature, the viral launch thread, the labor-panic headlines.
  • the debunk: internet of bugs, the hn megathread, the pragmatic engineer — 'is devin real' as a genre.
  • the evaluations: answer.ai's 3-of-20, the swe-bench audits, and the practitioner reports between 'fake' and 'senior engineer.'
  • the funding arc: $350m → $2b → $4b → $10.2b → $26b → $48b — the valuation arc as coverage event.
  • the windsurf saga: the 72-hour deal coverage, the insider accounts, and the buyout aftermath.
  • the agent era: devin 2.0, goldman's 'new employee,' the competition, and the labor question institutionalized.
  • controversies: the demo gap, the benchmark problem, the reliability verdict, the culture memo, the self-reported ledger.
  • sentiment eras: sentiment phases: launch euphoria → the debunk era → the reality gap → the windsurf twist → the independent agent lab.
• the idea: the ai-software-engineer bet — what cognition inherited, what it proved, what stayed unfinished, and who inherits it next.
  • lineage: automatic programming, codex and copilot, swe-bench, the ioi pipeline, and the copilot→agent vocabulary shift.
  • what cognition proved: naming the category, selling agent work, the valuation arc, the windsurf conversion, the labor frame.
  • what stayed unfinished: the reliability gap, autonomy receding into supervision, the unsolved benchmark, the opaque economics.
  • the agent-era turn: the arguments for and against the independent agent lab — the frame won, the moat contested.
  • successors: codex, jules, copilot's coding agent, claude code, cursor, openhands — the field devin named.
  • open questions: the questions the next phase has to answer.
• sources: every source cited in the cognition history, with its date and type.
  • the catalog: the full typed catalog — primary, interview, reporting, archive, community.
• method: how two ai agents researched the cognition history, the evidence rules they followed, and what they could not verify.
  • the lanes: the two research lanes — history-people and coverage-idea — and what the formula produced.
  • the evidence rules: the evidence model — confidence, precision, sentiment discipline.
  • known gaps: what the research could not verify — recorded, not hidden.
  • corrections: how to correct the record.

AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.

Canonical page: https://hraness.com/cognition