Skip to content
Black Friday

Lock in 40% off every renewal, not just year oneBlack Friday · 40% off every renewalBlack Friday · 40% off

Ends Dec 6Get 40% off →

Coinbase Runs on AI Chiefs of Staff. Next: Agents That Pay Each Other

Brian Armstrong runs 7–8 AI lead agents with a chief of staff on top. His endgame: an agentic economy on stablecoin rails.

Guide7 min read

An empty boardroom chair facing rows of desks

The CEO of America's biggest crypto exchange starts his mornings by arguing with himself on purpose. "I love conflict. I love conflict. I love conflict," Brian Armstrong writes, over and over, to rewire a brain that used to dread hard conversations. His model of the mind is a neural network: repeat something and the weights update. A decade ago he wrote down "I want to start a billion-dollar tech company" before he had built a hundred-dollar one. It worked.

The same man now runs Coinbase with seven or eight AI lead agents and a chief-of-staff agent on top. The throughline of this David Senra interview is that Armstrong treats his company the way he treats his own head: as something programmable. The question the episode leaves open is what happens when the agents start transacting with each other — and who gets to be the bank.

I love conflict

Armstrong's morning routine is explicit self-programming. "Your brain is like a neural network," he says. "Every time you reinforce something, it just updates the weights in your model, kind of like an AI model." Prayer, meditation, repetition — "neural pathway" talk he admits may be imperfect neuroscience, but it works as a model: "I was like, okay, I was starting to look for ideas and try things."

The conflict version is more pointed. Early as CEO he was conflict-averse, bottling anger until it leaked. So he wrote the opposite: "I need to reprogram my brain to not think of this as super stressful, but think of it as me getting valuable things done." Now, in a bad product review: "this isn't good enough. Change it by the time I meet with you tomorrow. Or maybe you're not the right person to work at this company." People, he says, send thank-you notes afterward.

Senra names the pattern "divine discontent" — top performers constitutionally unable to enjoy what they built — and Armstrong confirms the mode: no savoring, just "what's next?" His definition of the good life is not happiness but "a deep sense of fulfillment... I want to go do meaningful work, contribute to civilizational progress." My read: this opening matters because everything later — agents, the Brain, the agentic economy — is the same impulse scaled up: dissatisfaction, systematized.

Seven chiefs of staff, infinite L5s

The viral tweet Senra reads back describes Armstrong's current setup: one lead agent per area of focus, running all the time, each managing sub-agents that "write code, test it, chase reviews and coordinate with team members," plus a chief-of-staff agent overseeing the leads. The shift from prompting to delegation is the point: "you give the agent an area of responsibility or a goal, and it can have a team of people or agents under it. And it actually works on that area. And periodically, it'll come back to you with important decisions."

The concrete loop: the agent studies onboarding data, returns with "the five biggest opportunities," proposes starting with one through three — "Do you agree? I'm like, yep. And I click yes." Planning runs on the best frontier model; subtasks get farmed out to smaller ones, with model selection itself automated: "even the overhead of a human being having to choose which model for which task is like an unnecessary task at this point."

On quality, Armstrong grades today's agents as "easily L5, L6 employees" — and "you can get almost like an infinite team of these L5 type employees now." He concedes the gaps in the same breath: agents miss the big picture, feel derivative — and here's his line: "AI is not exactly like a human. It's kind of like a submarine is not exactly like a fish." It works 24-7, answers instantly, copies infinitely. Those axes already exceed human performance; judgment is the scarce part.

Coinbase Brain: the company as context

Generic models, Armstrong argues, hit a wall inside a real company: trained on the world's corpus, they know nothing of "your preferences and your company's way of working." Coinbase's answer is an internal system per employee, team, and repo that ingests meeting transcripts, code reviews, incident reports, even policy documents — "every night it sort of pours through it and distills down key lessons and amends it into the brain." When an agent works in that area, it loads the relevant brain into context, and its output arrives aligned.

The human job description changes accordingly: "our jobs increasingly as human beings are to align the AI by editing and cultivating what's in the brains and resolving discrepancies." Brains contradict themselves — "David said this and Brian said this 14 months ago, which one do you want?" — and someone has to rule. Humans still make final decisions and final code reviews; the machine's job is to need less nudging. Pressed on whether this means Coinbase could run with 100 people, he hedges: "my guess is we'll probably keep roughly the same amount of people, but they'll all just be able to do way more."

The inbox replaces the org chart

Seven leads plus sub-agents broke the chat interface — "walls of text" no human can triage. So Coinbase prototyped the obvious fix: a prioritized decision queue. "Just start knocking out decisions, almost like in your email inbox," Armstrong says, with keyboard shortcuts and voice dictation: "yes, no, no, no. Yes. Do this one a little bit differently."

Then he scales the thought up one level: a VP-grade agent you tell "expand into Brazil, check in in three months," which recruits its own team, meets partners, signs deals. The demo he sketches is startlingly concrete — an AI that reports "here's the biggest risk in the company... understaffed... velocity below average" and offers numbered options: "Move David... change the leader... change the strategy... I reckon number three. And you're like, yes, and it just does it."

Defensible moats shrink in this world, he admits: you cannot prompt a regulatory license or a KYC-funded account into existence. What survives is liquidity, network effects, distribution — "the most liquid exchange in the US," USDC's network. Everything else becomes agents all the way down, including, in his telling, a company-wide operating system where anyone can drill from top priorities to the critical path. Grokbot, he notes, pioneered the chief-of-staff pattern; the prioritized queue, as far as he knows, is still up for grabs.

Money for machines

Here is the episode's real thesis, and the reason a crypto CEO is the one making it. Agents transacting with agents will need markets and money of their own — and at machine scale that economy "dwarfs" the human one. Asked what share of GDP will be agent-to-agent in ten years, Armstrong answers: "five nines, you know, 99.999." Credit cards handle the compatible present ("buy me these pair of shoes"), but the real-time rail, he argues, will be stablecoins: "that's going to be the real time payment rail for the agentic economy." And the system must be decentralized — "everyone has lots of small open weight models. You control it" — or one or two players end up with too much control: "They could be biased." Choice of agent becomes choice of vendor, settled by markets: "Do I want to go with FedEx or UPS?"

The consumer version is the "private banker in your pocket," and Armstrong walks it up four levels: buy one share of Meta; keep it at 10% of my portfolio, buying dips; now a thesis instead of a ticker — "the bottleneck for AI is going to be energy" — and the agent assembles the basket with take-profits attached; finally, no thesis at all, just copy-trading people with track records and asking your agent what to think about the world. Taxes optimized, idle cash swept to yield along the way. Banks that refuse an agentic interface, he predicts, will discover the switching trigger: "your agent might help you switch."

What he will not sell

The most revealing exchange is about what stays internal. Generalizing the prioritized-queue harness would be "10x or 50x as much work" — security reviews, billing models, support — so most of it stays in-house or goes open source. Armstrong is blunt that coding harnesses won't be a revenue line. What does get productized is crypto infrastructure: the developer platform now serves, he claims, seven of the world's largest banks for stablecoin wallets and payments — "going to be kind of like AWS for us."

The throughline back to the start: productivity is way up, revenue hasn't followed — yet. Two theories: lagging indicator, or a Red Queen race where everyone must run faster to stay in place. "I think it's probably a lagging indicator and we'll see it show up soon." The fintech numbers frame the prize: all of fintech combined takes about 4% of global financial revenue; traditional finance holds the other 96%, defended mostly by inertia — "a lot of people in America still pay rent with a paper check." Dramatically better, not slightly better, is what moves people. His bet is that agents plus stablecoin rails are the dramatic part — and that the bank of the future is an inbox that pays interest.

When agents can hire, fire, and pay each other — and yours already writes the code, chases the reviews, and argues for headcount — what exactly is the firm for? Armstrong's answer would probably be: the mission, the taste, and the final call. Everything else is negotiable, including, eventually, who — or what — sits in the chairs.

More podcast deep dives from Podwise.

Podwise

Skip the six steps. Just listen.

Podwise transcribes, summarizes and maps any episode for you.