Cognition at $48B: coding agents move into your ops queue

Cognition closed over $2 billion at a $48 billion valuation today, led by Andreessen Horowitz and Accel. Four months ago the company raised at a $25 billion pre-money on a $492 million run rate. The valuation roughly doubled, and so did the revenue. But the number is not the story. The story is buried in the announcement: the coding agent is no longer sold as something you ask for a pull request. It is sold as something that wakes up when your monitoring does.

What actually happened

Cognition's own post puts run-rate revenue at "$492M to almost $900M" since the May round, and names the round: a16z and Accel leading, Founders Fund, General Catalyst and Avenir following, with 40-plus other backers including Benchmark, Kleiner Perkins, Greylock, NVIDIA and T. Rowe Price. Bloomberg reported the raise the same day. In May, TechCrunch reported $1 billion at $25 billion pre-money.

The product list is the interesting part. Cognition named Devin Auto-Triage for first-pass incident investigation, Devin Security Swarm for vulnerability triage, and Devin Automations, which start work from events in Slack, GitHub and Linear. Named customers: NVIDIA, GE Aerospace, Citi, Mercedes-Benz, Modal. The company also describes itself as an "independent agent lab" that picks and combines models rather than betting on one provider.

Why event-driven agents matter for your business

Event-triggered work is usage you did not schedule. A developer who opens an agent session decides to spend money. A PagerDuty alert at 2am does not ask. If you wire an agent to a webhook, put a hard cap on runs per hour and per repo before you wire it, not after the invoice.

The seat model stops describing the cost. We wrote in August about the gap between the round and the run rate. Four months later the run rate closed most of that gap, which tells you enterprises are paying per outcome on migrations and triage queues. Price your own pilot the same way: cost per closed ticket, not cost per developer.

"Independent agent lab" is a hedge they took and you probably did not. Cognition refuses single-model lock-in for itself. Copy that. Keep the model name and the agent vendor as config values, and keep an eval set of your own real tickets so you can measure a swap instead of guessing at one.

Start the agent on the queue nobody defends. Dependency bumps, flaky-test triage, log-noise cleanup. If the agent is wrong there, you lose an afternoon, not a customer.

Sources: Cognition, Bloomberg, TechCrunch.