Coinbase: AI now generates more than 95% of code

Coinbase reports that more than 95% of its code is now written by or with large language models, and that nearly 100% of its employees use AI tools on a daily basis. The company released the figures weeks after reducing its workforce by about 700 employees, roughly 14%, in May.

Rob Witoff, Coinbase’s head of platform, told employees and partners the share of AI-assisted code rose from an estimated 40% in February to between 95% and 100% today. CEO Brian Armstrong wrote in an internal email that AI has “dramatically” changed the pace of work and said the company intends to return to the speed and focus of its founding “with AI at our core.”

Witoff described a range of AI use across engineering tasks. Internal prototyping is now effectively fully automated. Core systems receive a mix of human and AI input. Cryptography and other security-sensitive code remain mainly led by human experts who review work one line at a time. AI is also used to test code, scan for vulnerabilities and verify mathematical correctness, with those processes involving more manual oversight than early-stage prototyping.

Teams have been reorganized around smaller, more senior groups. Witoff noted that two or three engineers now handle work that previously required teams of 10 or more. He reported that many junior development roles were affected in May and that reductions also reached marketing, legal, customer support and compliance functions.

Most Coinbase engineers run multiple AI agents at once, typically five to 10 per engineer. Witoff estimated AI agents are collectively performing the coding equivalent of about 1,200 employees today and projected agents could match the output of 100,000 employees by 2030. Coinbase said overall AI spending has remained flat even as use of AI-related tokens has grown.

Other crypto and technology firms have cited AI-related efficiency gains in workforce reductions earlier this year. Coinbase executives say they will maintain human oversight for high-risk and security-critical systems while expanding automated development and testing for lower-risk and experimental work.