Coinbase's Brian Armstrong runs his work through a network of AI agents

Coinbase chief executive Brian Armstrong said on November 7 that his working style has changed over the past few weeks: instead of handling tasks himself, he now routes work through a standing network of AI agents.

In a post on X, he described one lead agent for each of his areas of focus, running continuously, with a chief of staff agent overseeing them. Each lead in turn manages its own sub-agents, which write code, test it, chase reviews and coordinate with team members.

The agents funnel recommendations and pending reviews into a single queue. Armstrong reviews changes through a local web app that shows screenshots and video clips, and can work through several at a time using voice commands or keyboard shortcuts. Human review and security checks remain the gate before anything is approved.

The system also learns from his past calls. Armstrong refers to it as a Brian brain: the agents weigh his earlier decisions when proposing next steps, so that future work tracks his expectations more closely.

The update covers his personal workflow rather than a company-wide rollout, a distinction Coin Edition makes explicit in its report.

Still, it fits the direction Coinbase has been pushing. In October, chief policy officer Faryar Shirzad discussed research into automated payments that need no human involvement, identifying stablecoins as a fit for frequent, low-cost machine transactions and pointing to Coinbase's x402 protocol, which lets servers and AI agents exchange payment requirements and settle transactions.

Armstrong has separately argued that stablecoins could become a standard payment rail for AI agents. Once agents take on productive work such as writing and testing code, questions about who pays an agent and who is accountable for what it does stop being thought experiments.

The pattern worth watching is whether the layered structure of lead agents, sub-agents and a human approver spreads from one executive's experiment into how organisations are run. It also exposes the practical constraint: the final call still sits with a person, which makes review and security capacity, not code generation, the bottleneck.