Coinbase Says AI Now Writes 95-100 Percent of Its Code

Coinbase, one of the world's largest cryptocurrency exchanges, has acknowledged that nearly all of its software is now written with the involvement of artificial intelligence. Rob Witoff, the company's head of platform, said the share of code created by or with the help of large language models has reached 95 to 100 percent, up from 40 percent as recently as February 2026.

The jump from 40 to nearly 100 percent within a few months is one of the most dramatic claims about coding automation made by a major tech company to date. Earlier, in September 2025, Coinbase CEO Brian Armstrong wrote on X that about 40 percent of the code produced daily at the company was generated by AI, and set a goal of surpassing 50 percent by October. According to the latest statements, reality has outpaced those plans by several months.

The math of digital workers

The key element of Witoff's statement is the conversion of AI agent work into full-time positions. Most Coinbase engineers run 5 to 10 agents at once, which write, test and review code. The combined output of these agents, calculating compute time as the equivalent of a 40-to-60-hour work week, is said to match the work of about 1,200 full-time employees.

Witoff went further, offering a forecast for the end of the decade: by 2030, AI agents at Coinbase could be doing work comparable to employing 100,000 people. That figure dwarfs the company's current headcount, which stands at just over four thousand employees following the May layoffs.

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 - Rob Witoff, Head of Platform, Coinbase

Layoffs and a new team structure

The claims about code automation are not an isolated press statement but a continuation of changes Coinbase had already put in place. In May 2026, the company laid off around 700 people, or 14 percent of its workforce, with management directly linking the decision to the growing role of artificial intelligence in day-to-day engineering work. Coinbase's traditional project team, made up of one product manager, one designer and eight engineers, has been replaced by smaller groups: two to four people now work alongside roughly ten AI agents that act like team members on Slack channels.

Agents independently open pull requests and prepare draft solutions, which people then review and approve. The result is said to be a doubling of code shipped per developer year over year, with the company's top engineers sending as many as around 100 pull requests a week. Coinbase stresses that despite the rise in code volume, the number of bugs and incidents per line of code has fallen.

Where automation stops

The company notes that the level of automation is not uniform across all areas. Cryptographic code, which is critical to security, remains largely under human control, with experienced cryptographers reviewing it line by line. At the other extreme are internal prototypes, which the company says are now built fully automatically. Production systems fall somewhere in the middle, with AI heavily used for testing, vulnerability checks and mathematical verification, though the final call on what makes it into the repository still rests with people.

The 95-100 percent figure itself also raises questions of interpretation. Describing code as written with AI can cover anything from simple autocomplete suggestions in an editor to fully autonomous pull requests prepared without any human involvement. Witoff folds both extremes into a single number, which makes it hard to compare Coinbase's claims with similar statistics published by other tech companies.

The compute cost calculation

The scale of the AI agent rollout also raises questions about infrastructure costs. Armstrong has pointed out that open-source models can be up to 99 percent cheaper to run at inference than the most advanced closed models, even though they lag them by three to six months in capability. Coinbase expects that over the next 12 to 18 months, as much as 80 percent of the company's compute workloads will shift from expensive frontier models to cheaper alternatives, once the quality gap stops mattering for routine tasks.

For the Polish market, where AI is only slowly making its way into the daily work of development teams, the Coinbase case points to the direction large tech organizations may follow in the coming years. Companies evaluating AI deployments will have to grapple with similar questions: how to translate agent output into full-time equivalents, where to draw the line on automation, and how to restructure teams that are no longer counted mainly in people.