AI-assisted development, the practice of using machine-generated suggestions to help engineers write and review software, now accounts for more than 95% of the code shipped at Coinbase. Rob Witoff, an executive at the crypto exchange, described the company as leaning harder on AI for execution tasks. The carve-out: judgment and strategy still belong to people.
What that 95% figure does and does not say
A number like "95% of code written with AI help" is a wide tent. At one end of the range, a developer accepts an autocomplete suggestion on a single line. At the other, an AI agent drafts entire modules and a human reviews the output. Witoff did not specify where Coinbase's workflow sits on that spectrum, so the figure describes participation, not authorship.
That gap matters. Knowing AI touched the code tells you nothing about how much of the final result the AI actually produced. Anyone reading this as a productivity multiplier or a headcount signal should note what the disclosure does not say.
The "high-agency human" frame
Witoff's comments centered on a phrase worth unpacking: "high-agency humans." High agency, in practice, means the capacity to take initiative, weigh competing risks, and make consequential calls without a step-by-step script. Coinbase's stated position is that AI handles well-defined, repeatable tasks. The employees who own strategy and bear responsibility for outcomes that carry real weight still need to be human.
A figure above 95% means AI involvement has become the default in Coinbase's engineering workflow, not the exception. Whether the boundary between human judgment and AI execution holds as models improve is the question Witoff's comments left open. For now, Coinbase's public position is that execution can be automated. Judgment cannot.