Brian Armstrong has reorganized his workday, and most of the new org chart is software.

In a post on X on Saturday, the Coinbase co-founder and CEO said he now has "one lead agent for each area of focus, running all the time, and a chief of staff agent that oversees them." Each lead, he wrote, "manages its own sub-agents that write code, test it, chase reviews, and coordinate with team members."

Lots of executives talk about using AI. Armstrong's post stands out because of what he says he handed off: the tracking, chasing and coordinating that usually falls to a chief of staff or a layer of middle managers. It's a rare, detailed look at a public-company CEO running his own calendar this way.

Everything lands in one queue

All of that work ends up in one place. "Everything I need to decide or review comes into one prioritized queue, a simple local web app," Armstrong wrote. Each product review arrives with screenshots and a video demo of the feature, and he clears the queue in batches, "mostly by voice and keyboard shortcuts."

It's the least flashy part of the post and probably the most useful. Anyone running more than a few coding agents knows the problem. Each one lives in its own chat window and wants attention, so the hours saved on the work get spent checking on it. Armstrong said that working from one queue, "instead of switching between dozens of chats, has cut a lot of mental overhead."

The traffic isn't all one way, either. He wrote that the leads "bring back ideas I hadn't considered," which suggests they're doing some scoping on their own instead of just closing tickets.

He's trying to teach the system his judgment

The bigger claim is about judgment. Armstrong said the system "keeps learning my preferences and distilling my judgment into a 'Brian brain,'" so its recommendations and first drafts improve before they reach him. "The goal is fewer corrections per change," he wrote.

A good human chief of staff does the same thing over time. After a few months they know which memo will come back marked up, and why. Armstrong is trying to get software to that point, with every approval and rejection in the queue serving as another lesson.

He didn't say how the Brian brain is built, which models it runs on or where his preference data lives. At a company that holds customer funds, someone will want those answers.

Humans still sign off on every change

Armstrong drew the line plainly. "Every change still goes through the same human review and security checks," he wrote. "It is eliminating tasks, not human judgment."

That caveat carries more weight at Coinbase than at most companies. It's a publicly traded exchange that answers to financial regulators, and the agents he describes write and test code. A setup where agents shipped changes on their own would be a different story. His version keeps people as the approval step and uses agents to get work there faster.

It's also, by his own description, a personal workflow. The post doesn't say the setup is being rolled out across Coinbase.

Coinbase has been building toward this all year

The post fits a pattern. In April, Coinbase began testing internal agents named "Fred" and "Balaji," modeled on co-founder Fred Ehrsam and former CTO Balaji Srinivasan, that show up in Slack and email like any other colleague. Armstrong said at the time that he expected Coinbase to have more agents than human employees "at some point soon."

On the All-In podcast, he has also described an internal agent he calls the "Oracle of Coinbase," connected to the company's Slack messages, Google Docs and Salesforce data. He said he uses it to ask what he might be missing, rather than to give it orders.

The business strategy points the same way. Coinbase launched Agentic Wallets in February so AI agents can hold funds and transact onchain, and it backs the x402 protocol for machine-to-machine payments. Armstrong closed Saturday's post by tying his own experiment to that push, calling it an exciting time "especially as we build the economy for agents (agentic finance)."

Our take

We think the queue matters more than the agents. Spinning up a dozen agents is easy now. Keeping up with what they produce is where most people stall, and Armstrong's fix of one prioritized list, with demos attached and cleared in batches by voice, is practical and easy to copy.

The Brian brain is the piece to watch. If it works, more of a CEO's job becomes training the system that drafts his decisions. If it doesn't, he's reviewing more work than before, only faster. He hasn't shared any numbers on corrections per change, the metric he says he cares about most.

Founders who want to try this should start small. Build the queue before adding more agents. Attach proof to every item so review doesn't turn into redoing the work. Save each correction somewhere the system can use it again. Keep a person on the approval step for anything that touches money or production code.

Armstrong told readers to "keep pushing in this direction." The thing to check in six months is how much of the judgment in that queue is still his.