Coinbase says over 95% of its code is now written with AI
Coinbase head of platform Rob Witoff says between 95% and 100% of the crypto exchange's code is now written by or with large language models. The company is leaning more on AI for execution, but Witoff emphasizes it still needs high-agency humans for judgment and strategy—showing how far over Coinbase's code now flows through AI before it ships.
The disclosure, reported by Cointelegraph, offers a concrete look at how one of the largest U.S. crypto platforms is reorganizing engineering around automation. Witoff told the outlet that effectively 100% of Coinbase employees use AI daily, while code output sits in a wide spectrum between human-led and fully automated workflows.
Key Takeaways
- Coinbase estimates 95% to 100% of its code is written by or with LLMs, up sharply from a 40% figure cited earlier in 2026.
- Internal prototyping is now effectively fully automated, while core cryptography and security testing still rely heavily on human oversight.
- Most engineers run five to 10 AI agents at once; Witoff said those agents collectively perform work equivalent to about 1,200 employees.
- Smaller senior teams—often two or three people—now handle projects that previously required 10 or more staff.
- Coinbase frames the shift as execution support, not a replacement for human judgment on strategy and critical systems.
How did Coinbase reach 95% AI-written code?
Witoff said the current figure is more than double Coinbase's February estimate, when the company reported that 40% of its code was written with AI. That acceleration mirrors broader adoption across tech and crypto, but Coinbase's numbers stand out for how quickly the share climbed within a single year.
He described a spectrum of use cases rather than one uniform process. Internal prototypes are now effectively 100% automated, while core systems sit in the middle. Core cryptography, by contrast, still depends mainly on human input.
Where does human judgment still matter?
Despite the headline figure, Witoff stressed that Coinbase still needs what he called "high-agency humans" for judgment and strategy. AI helps with execution, but humans remain central where mistakes carry outsized risk.
Coinbase also uses AI to test code, check for vulnerabilities, and verify math—but Witoff said that verification work is far more manual than building internal prototypes. The company is not treating every line of code as equally safe to generate without review.
Why does this matter for crypto and fintech?
The shift has let Coinbase reorganize around smaller, more senior engineering teams. Witoff said two or three employees can now handle workloads that previously required 10 or more people, with most engineers running five to 10 AI agents at any given time.
For investors and users watching the exchange, the story is partly about efficiency and partly about trust. A platform that moves most code through LLMs must still prove that security-critical paths—custody, trading infrastructure, and cryptography—remain rigorously checked. Follow more Fintech & Crypto Alerts for updates on how exchanges and startups are deploying AI across financial infrastructure.