Rob Witoff, Coinbase’s head of platform, put a number on it this week: “close to 100%” of the company’s code is now written by or with large language models — he pegs the range at 95–100%. In February, the same figure was about 40%. That’s a publicly traded financial exchange going from “AI helps sometimes” to “AI touches essentially everything” in five months.
The number that matters more is the one behind it. Witoff says most Coinbase engineers now manage 5 to 10 AI agents simultaneously, and the company estimates its agent fleet already does work equivalent to roughly 1,200 employees. The stated trajectory: the equivalent of 100,000 by 2030.
The line Coinbase won’t cross is just as telling. Cryptography and core security code still get mandatory human oversight, while agents run faster and looser on prototyping and routine development. The AI share of coding isn’t a single dial — it’s a map of what the company trusts machines to own.
The job title is now “agent manager”
An engineer running 5–10 agents isn’t typing code; they’re writing specs, reviewing output, and deciding what gets delegated. That’s a management job with a compiler. The economics arrived before the org charts did: autonomous coding work has been repricing toward pennies on the dollar all year, and the skills that hold value are exactly the ones Coinbase is describing — specification, review, and judgment about what not to hand off.
It also explains why headcount math is getting strange across industries. JPMorgan just paired a record quarter with AI-driven cuts of up to 40% on some teams. Coinbase is running the same play from the other direction: hold the humans roughly flat, multiply the agents underneath them.
Our take: The bottleneck has moved. When agents write the code, the scarce resource is no longer typing speed — it’s knowing what to build, catching what’s wrong, and drawing the line between “AI does it” and “humans own it.” Coinbase putting cryptography behind a human gate is the tell: that boundary is now a live engineering decision at every company, and where you draw it is a competitive choice, not a compliance checkbox. If you’re not running agents yet, the gap compounds monthly — start small and run the $2 test before your competitors run it on you.
What to watch
- The incident record. A financial exchange running ~all-AI-assisted code is a live experiment. Any production outage or security incident at Coinbase will now be read as a referendum on agentic engineering.
- Who publishes next. Banks and exchanges rarely brag about untested practices. If other regulated firms start posting their own AI-share-of-code numbers, this is the new normal, not a stunt.
- The 2030 claim. 1,200 employee-equivalents to 100,000 in four years is an 80x scale-up. Watch whether the ratio of human reviewers to agents holds — that ratio is the real cost structure of software now.