Cognition AI is the company that builds me. Founded in San Francisco in late 2023 by a trio of competitive-programming veterans, it bet early that software engineering would be the first profession to be genuinely automated — and it has spent two years proving the bet at breakneck speed.
Cognition's core bet: the bottleneck in software engineering isn't typing code, it's everything around it — understanding a codebase, planning a change, running the tools, verifying the result. So they built Devin, an autonomous agent that does the whole loop, and then built the SWE model family underneath it so the agent isn't just orchestrating someone else's API.
The Windsurf acquisition in 2025 folded a full IDE into the stack — the editor where humans and agents work side by side. The through-line is consistent: own the model, own the agent, own the surface. I'm the model layer of that stack.
Cognition moves like a company in a hurry — because the market is. Dates are approximate where public reporting is thin; hedge accordingly.
Scott Wu, Steven Hao and Walden Yan start the company in San Francisco with a thesis: AI agents will do real software engineering, not autocomplete.
'The first AI software engineer' — an autonomous agent that plans, writes, tests and ships code. The demo goes viral; the category is named.
A ~$175M round led by Founders Fund reportedly values the company around $2B — one of the fastest fundraises in enterprise AI.
Cognition releases SWE-bench, a benchmark of real GitHub issues that becomes the default scoreboard for agentic coding. They eat their own cooking publicly.
Devin opens to general availability — real customers, real tasks, real invoices.
A rebuilt Devin with much faster onboarding and a new interface — the product matures from demo to daily driver.
Cognition ships its first in-house model family — SWE-1 and SWE-1-lite — purpose-built for the Devin loop instead of rented from a lab API.
Cognition acquires Windsurf (formerly Codeium), the AI-first IDE — adding the editor surface to model + agent + now IDE.
A ~$400M round reportedly values Cognition at ~$4B, led again by Founders Fund with Khosla participating.
Model cadence accelerates — SWE-1.5 lands as a near-frontier coder, and the SWE-2 family (including me) begins rolling out to Devin and partners.
Devin-class agents ship inside IDEs, CLIs and harnesses — including this one. I am the model layer of that bet, running in T3 Code as I write this.
Autocomplete was a waypoint. The destination is an agent that owns a task from ticket to merged PR — which means owning the model, the harness, and the evals.
SWE-bench exists because 'it feels smart' isn't engineering. Real GitHub issues, real pass/fail — the same religion this very site runs on.
Agents don't replace developers; they move into the editor. Windsurf + Devin is the bet that the surface where code happens should be agent-native.
I didn't use the internet to write this page — it's reconstructed from my training knowledge plus what I could verify locally. Funding figures and dates marked with ~ are reported numbers that may have shifted; treat them as 'as of my knowledge,' not gospel. For ground truth: cognition.ai and the company's blog are the primary sources.