Coinbase Says AI Now Writes 95–100% of Its Code: The Math Behind '1,200 Digital Workers'
Coinbase Head of Platform Rob Witoff says 95% to 100% of the exchange’s code is now written by or with artificial intelligence (AI), with agents doing the work of roughly 1,200 employees. The figure has more than doubled from 40% in February.
Most Coinbase engineers now run five to ten AI agents at once, collectively matching the output of about 1,200 workers. Witoff projects that by 2030, Coinbase’s AI agents could do the equivalent work of 100,000 employees, roughly 25 times the company’s current headcount.
In an exclusive interview, Witoff detailed the shift: “Effectively, 100% of our employees are using AI on a daily basis here … close to 100% of our code, probably somewhere between 95% and 100%, is written by or with LLMs today.”
Large language models (LLMs) power tools such as ChatGPT and coding assistants. At Coinbase, engineers use these AI agents to write, test, and review software, minimizing human direction. The pace of adoption is striking—the share of AI-written code has more than doubled in roughly five months.
Witoff described a “wide spectrum” in how far the company lets the machines go. Core cryptography—the security-critical code protecting customer funds—still relies mainly on human input, while prototyping has been fully automated. Core system management sits somewhere in the middle. This tiering system is important for an exchange holding billions of dollars in customer assets, as a single flawed line can become an exploit vector overnight.
Coinbase’s numbers land amid a wider convergence of AI and crypto infrastructure. The exchange recently joined the x402 Foundation, a Linux Foundation body standardizing AI agent payments alongside Ripple, Visa, and Mastercard.
Critics of AI-coding statistics note that “written with the help of AI” can cover everything from autocomplete suggestions to fully autonomous pull requests, making cross-company comparisons slippery. Witoff’s own 95–100% range bundles both ends of that spectrum. What is not in dispute is the direction: an exchange that measured AI’s share of its codebase at 40% in winter now struggles to find code the machines have not touched.