📰 Key Highlights

Coinbase now has over 95% of its code written with AI assistance, more than doubling the 40% figure announced back in February this year — a clear sign that its AI adoption pace is rapidly accelerating. The number was disclosed by Coinbase platform lead Rob Witoff to Cointelegraph, and it comes shortly after the company laid off 700 employees (14% of its workforce) earlier this year. CEO Brian Armstrong wrote in a memo to staff that AI has “dramatically” changed the pace of work, and that the company needs to “get back to startup speed and focus, with AI at the core.”

Witoff pointed out that the level of AI usage varies hugely depending on the task: writing core crypto algorithms is still primarily human-driven, with top cryptographers in the industry going through every line carefully; AI is instead heavily used to test whether code runs correctly, check for vulnerabilities, and verify mathematical correctness — these are the more labor-intensive parts. By contrast, internal prototype development is now nearly 100% automated.

This shift allowed Coinbase to reorganize into leaner, more senior-heavy teams. Work that previously required 10+ people can now be handled by just 2 to 3 employees. The May layoffs hit junior developer roles the hardest, but the scope also extended into marketing, legal, customer support, and compliance. Witoff said most Coinbase engineers currently run 5 to 10 AI agents simultaneously, and the work done by these agents combined equals the output of roughly 1,200 employees. He also projected that by 2030, AI agents could be doing work equivalent to 100,000 employees. Despite the growth in agent usage, Coinbase’s AI spending has remained “flat.”

💬 JudyAI Lab Take

Coinbase’s latest disclosed figure — over 95% of its code is AI-assisted, more than doubling from 40% in February — is a growth rate that every AI builder should pay close attention to.

This case reflects a clear division of labor: AI isn’t wholesale replacing engineers. Instead, it’s splitting work by task type with precision. The high-trust, high-scrutiny parts — like core crypto algorithms — are still being guarded line by line by top cryptographers, while AI handles the repetitive but critical tasks: testing, vulnerability checks, and math verification. Internal prototyping is essentially fully automated. This “humans guard the last line of defense, AI handles validation and prototyping” setup shrank teams from 10+ to just 2–3 people for the same workload. Witoff mentioned that engineers routinely run 5 to 10 AI agents in parallel — this multi-agent parallel work pattern might be an important reference for how teams will be structured in the future.

For you as a reader, here’s something worth thinking about: in your own workflow, which steps are “verifiable but repetitive” — the kind you could hand off to AI agents first?

📅 Source Info

🔗 Further Reading