Brian Armstrong now runs Coinbase with seven or eight lead AI agents, a chief of staff agent, and an internal “Coinbase Brain” that turns every team’s meetings, incidents and code reviews into memory overnight. In this hour with David Senra, the Coinbase co-founder and CEO explains how that setup works, why he thinks almost every transaction will soon be agent to agent, why productivity gains have not yet shown up in revenue, and why traditional finance still holds 96% of the market.
TLDW
Armstrong stopped prompting AI and started giving agents jobs: an area of responsibility, sub-agents underneath, and a prioritized inbox of decisions he clears with keyboard shortcuts and voice dictation. Coinbase Brain feeds each agent the accumulated knowledge of the team or code repository it works in, and the human job becomes editing that knowledge and settling its contradictions. Fine-tuned open-weight models beat frontier models on narrow work like compliance reports. He expects agent-to-agent commerce to dwarf the human economy and run on stablecoin rails, wants AI decentralized the way crypto decentralized money, and is cutting third-party vendors as cyber incidents roughly double each year. He admits AI has not materially moved Coinbase revenue yet, argues the real prize is the 96% of financial services revenue fintech has never touched, lays out four levels of AI investing that end with a private banker in your pocket, and says his biotech New Limit will not use humanoid robots until one can make an espresso.
Thoughts
The most honest thing Armstrong says is that AI has produced a “very material” jump in productivity at Coinbase and has not materially moved revenue. He offers two explanations, a lag that should close in about six months or a Red Queen race where everyone gets faster and nobody gains ground, and he picks the lag because traditional finance is not adopting AI as fast as fintech. I think he gives the Red Queen too little weight, and his own argument earlier in the hour is the reason. He says software is getting easy to prompt into existence while regulatory licenses, KYC’d funded accounts, exchange liquidity and USDC’s network effects are not. If code gets cheap for everyone, the gains flow to customers and to whoever owns the things code cannot make. That is good news for Coinbase, which owns plenty of those things, but it suggests the payoff shows up as defended margin and faster shipping rather than as a revenue step change.
His answer on whether to sell Coinbase’s internal AI tools is the same thesis from the other side. Generalizing a tool built for one company is “10x or 50x as much work” once you add every customer’s security and billing needs, and he expects coding harnesses to end up as open-source projects that every company runs itself, so “a huge part of the value chain might not get captured.” Yet Coinbase happily sells its crypto infrastructure through the Coinbase Developer Platform, where he says seven of the world’s largest banks are building stablecoin wallets and payments, and which he compares to AWS. The line he draws is clear: sell what needs custody, licenses and blockchain connections, and keep the AI harness, because anyone can rebuild that.
Coinbase Brain is the most concrete picture of AI-era management in the episode. Every employee, team and GitHub repository has a brain that ingests A/B tests, user studies, outage reports, meeting transcripts, code reviews and accounting policy, and distills lessons from them each night. The human job, in Armstrong’s words, is to align the AI “by editing and uh cultivating what’s in the brains and resolving discrepancies,” such as a call David made last week that contradicts one Brian made 14 months ago. That reframes what a manager is for: keeping the company’s memory consistent so agents can do the work in one pass. It also explains why he expects Coinbase to keep roughly its 4,000 people rather than shrink to 100. Someone has to own the memory, and humans still make the final decisions and code reviews.
The fintech argument has a sharp point inside it. Armstrong says a slightly better product never moves people off their bank, the way nobody left Google for a slightly better search engine; it took something dramatically different. His bet is that agents are that difference. If legacy banks will not let a customer’s agent operate the account fully, that alone is a reason to switch, and “your agent might help you switch.” His four levels of AI investing (place the order, build a position over time, turn a thesis into a basket, then ask a smart friend) are a useful ladder. The trust problem sits at levels three and four, which is presumably why he stresses strict controls and keeping the customer in charge.
Key Takeaways
- Armstrong changes his own habits by writing a line down over and over. In his 20s it was starting a billion-dollar tech company, which he did within six or seven years of moving to Silicon Valley. As a conflict-averse new CEO it was “I love conflict.”
- He aims for fulfillment, not happiness. Happiness is a fleeting high; fulfillment comes from meaningful work and progress over a long arc, and he thinks it requires doing hard things.
- He keeps a printout of the Civilization tech tree in his home office and says he wanted to “add a leaf” to it. The game also taught him that there are many ways to win, not just military conquest.
- Coinbase moved from prompting AI to giving agents an area of responsibility. The agent studies the data, proposes the top opportunities and comes back for decisions.
- He rates today’s agents as L5 or L6 employees in tech-company leveling, available in effectively unlimited numbers, while noting they can still miss the big picture.
- Small open-weight models fine-tuned on internal data now beat frontier models on narrow tasks at Coinbase, such as drafting compliance reports, and run faster and cheaper. Most employees use an “auto” setting that picks the model for them.
- Coinbase Brain keeps a brain for every employee, team and code repository, updated nightly, and agents load it into context before they work.
- His personal setup is seven or eight lead agents, each with three to five sub-agents, and a chief of staff agent. Decisions arrive in a prioritized queue so five spare minutes go to the choices that unblock the most work.
- Asked what share of transactions will be agent to agent in 10 years, he answered “five nines,” 99.999%, and expects stablecoins to be the real-time payment rail for that economy, with credit cards kept for backward compatibility.
- He wants AI decentralized, with lots of small open-weight models people control, rather than one or two big players with too much control.
- Cyber incidents across crypto and traditional finance have roughly doubled year over year, he says, so Coinbase is rebuilding vendor tools in house. Frontier agents review every line of code before production, and 99% of the crypto it holds is stored offline.
- He calls the precautionary principle “super dangerous,” citing the Nuclear Regulatory Commission and past panics over bicycles, violent video games and civilian cryptography.
- All fintech combined captures about 4% of global financial services revenue. The opportunity is taking share from traditional finance, plus an agentic economy whose size nobody knows.
- Humanoid robots are the next step change, he says, but New Limit, the biotech he co-founded, will not test them in the lab until they can work as baristas.
Chapters
00:29 Writing It Down Until It Sticks
Armstrong does not treat written goals as manifesting. Repetition keeps a goal top of mind, so you notice chances to act on it, the way training updates a neural network’s weights. He wrote down that he would start a billion-dollar company before he had started any company, and later wrote “I love conflict” until hard conversations stopped feeling like stress. Now when a product review falls short, he tries to turn the anger into what he calls empowered anger, and staff sometimes thank him afterward for saying it.
04:50 Divine Discontent and the Civilization Tech Tree
He never feels the Coinbase app is good enough and pings 30 or 40 people a day about it. Senra connects this to the divine discontent that David Ogilvy, author of Ogilvy on Advertising, saw in great creators. Armstrong rarely stops to celebrate, saying the goal is fulfillment over the long arc, and that life should be hard like a good video game. Civilization and StarCraft scratched the same itch as business, only over 20 minutes instead of decades.
17:20 Giving Agents a Job and Building Coinbase Brain
Agents at Coinbase get goals, not prompts. One assigned to improve international onboarding comes back with the five biggest opportunities and proposes starting with three. Planning uses the most expensive frontier models, which farm subtasks out to cheaper ones. Armstrong compares AI to a submarine next to a fish: superhuman in some ways, short of a person in others. Coinbase Brain distills each team’s history nightly so agents can get the work right with less correction.
24:45 A Prioritized Inbox and an Operating System for the Company
Walls of chat text overwhelmed him once he had many agents running, so Coinbase prototyped a queue of decisions with buttons always in the same place. He is also sketching a company-wide view of priorities, projects and critical paths, where the AI could flag a dependency that blocks 14 downstream projects and propose four fixes. The next layer up would be a VP-level agent told to expand into Brazil. Software is easy to prompt into existence; regulatory licenses are not.
29:18 An Economy of Agents
People, then teams, companies and countries all achieve more together, and agents will need their own markets, money and specialists to cooperate across companies. Armstrong expects 99.999% of transactions to be agent to agent within 10 years and agrees with Senra that an economy with effectively a trillion workers would be far larger than today’s. He says crypto and AI are better together, and that AI should be decentralized so no one or two players control it.
35:05 Fewer Vendors, Tighter Security
The agent result that surprised him most involved replacing a vendor he would not name. People in the industry called October “Black October” for the pace of breaches, and Armstrong decided third-party vendors were one of Coinbase’s biggest risks, so the company is rebuilding their tools in house. He says he is so energized by building right now that he has to tell himself to go to sleep.
38:14 Why People Fear AI, and the Precautionary Principle
Senra asks why so many people use AI and still hate it. Armstrong’s answer is evolutionary: people who feared change lived longer, and society needs only a few who welcome it. He admits AI labs’ messaging could have been better but thinks the fear runs deeper. People can use Amazon and worry about it at the same time. He walks through past panics over bicycles, video games and cryptography, and argues regulators should wait to see the real risk before writing rules.
44:50 Keep the Tools, and the Revenue Paradox
Shopify has turned down requests to sell its internal AI tools, while Ramp has decided to sell its own. Coinbase has decided to keep its own for now, because generalizing them is far more work and coding harnesses will likely be open source. It does sell crypto infrastructure through its developer platform. Armstrong says he is trying not to launch many more products, and that AI’s productivity gains have not yet shown up in revenue, which he expects to change.
49:33 The 96% and a Private Banker in Your Pocket
Traditional finance keeps about 96% of global financial services revenue because the world has inertia; plenty of Americans still pay rent by paper check. Armstrong thinks crypto as the rails and AI as the interface can change that. His four levels run from placing a single order, to building a position over time, to turning a thesis, such as energy and data centers being the bottleneck for AI, into a basket, to asking an agent or a community leaderboard what to do. AI should also move idle cash and save on taxes automatically.
56:47 Robots, Espresso and an Agent Called Retail App Quality
What AI still cannot do for him is act in the physical world, and humanoid robots building factories or running labs would be the next big shift. New Limit will wait until a robot can do the barista job. One of his lead agents is named Retail App Quality. It suggested an X post asking customers for complaints, summarized the replies into the top five issues, and put five worker agents on them. Armstrong calls the arrangement a collaboration: sometimes he pushes the agent, sometimes it brings ideas to him.
Notable Quotes
“I love conflict. I love conflict. I love conflict”
Brian Armstrong, on the line he wrote over and over to get past conflict aversion as a young CEO
“AI is not exactly like a human. It’s kind of like a submarine is not exactly like a fish”
Brian Armstrong, on judging agents against human employees
“software is getting easier and easier to prompt into existence, but like regulatory licenses are not.”
Brian Armstrong, on which parts of Coinbase stay defensible
“I’m just so excited I can’t sleep at night.”
Brian Armstrong, on building in the current moment
“I think the precautionary principle is super dangerous.”
Brian Armstrong, on regulating new technology before its real risks are known
“it’s like everybody now has a private banker in their pocket”
Brian Armstrong, on the fourth level of AI-assisted investing
“can it make you an espresso? It can’t do barista yet.”
Brian Armstrong, on why New Limit is not bringing humanoid robots into its lab
Watch the full conversation here.
Related Reading
- Brian Armstrong, background on Coinbase’s co-founder and CEO.
- David Senra’s show notes for the episode, with links to the tweets read on air.
- Precautionary principle, the regulatory idea Armstrong calls super dangerous.
- Red Queen hypothesis, the running-to-stand-still idea behind his second theory of why revenue lags.
- New Limit, the epigenetic reprogramming company Armstrong co-founded.