An Account for Every AI Agent: Credit Card Fees Can't Carry the Machine Economy
Coinbase wants every AI agent to have its own account: 76% of machine payments are under 30 cents, and 30 cents is exactly the credit card minimum fee — the old payment rails can't carry the machine economy.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
76% of machine payments are under 30 cents
Armstrong says agents are already generating a flood of transactions worth a few cents: VC firms scraping paywalled data for research, recruiters crawling LinkedIn, one agent calling another specialised agent as a tool to fetch source data. But the minimum per-transaction fee on a debit or credit card is a flat 30 cents plus a percentage, so the old rails basically don't work below $1 — paying 30 cents to move 1 cent makes no sense. The number he gives: roughly 76% of the agentic commerce transactions they see are under 30 cents. That is the hardest argument for stablecoins and crypto rails cutting into the machine economy.
— Brian ArmstrongBuild a "brain" for every service
Coinbase has built a "brain" for every team, every repo, every service and even every employee: it holds every incident that service has ever had, the financial controls that must be enforced, every A/B test ever run, the full GitHub history, and which PRs were accepted or rejected. When an agent comes to change code, it reads this brain first. The truly critical step: when a human reviewer finds the agent missed something, they are not allowed to just hand-edit and merge — that correction must be written back into the brain. That way one correction covers all future scenarios, and the one-shot PR rate keeps climbing, something Armstrong calls "recursive self-improvement".
— Brian ArmstrongThe CEO runs 10 agents on code at once
Armstrong says that now, when he hits an idea or a bug, he no longer pings the team all day; he drops the comment into an internal harness (they call it Toshi) and sends over a finished PR for review — he says it's addictive. He has also pushed the CEO back into writing features: have an expensive frontier model break a complex feature into three phases of ten items each, then have it recommend which cheaper model to use for each subtask, then run 10 agents in parallel on phase one. He says that over the course of recording this podcast, all 10 items got done and were waiting for his review.
— Brian ArmstrongWhat gets killed is tasks, not people
On whether AI will make companies smaller, Armstrong's judgment is: productivity really is rising, but that's a separate matter — what gets automated or eliminated is tasks, not people. He doesn't think a company like Coinbase shrinks to 10 people; rather, the people already there can do vastly more and everything speeds up; and yes, there will be two-person and five-person companies doing volumes that used to be unimaginable. The result is that the cost curve of revenue relative to headcount changes: the same group of people produces far more, and margins should rise over the long run. He doesn't close the door on it — some companies will still hire at the same pace.
— Brian Armstrong88% of revenue no longer comes from Bitcoin trading
Armstrong cites the numbers from the earnings reports: 88% of revenue now comes from non-Bitcoin trading. The interesting part is that Coinbase's stock price is still highly correlated with the Bitcoin price. He says he loves Bitcoin, sees it as digital gold and sound money for everyone in the world, but crypto is updating every part of the financial system. The tokenised equity product he just announced is currently only available outside the US, and they are talking to regulators about a US path; he stresses these are real equities — backed 1:1 by real securities held in custody, not synthetics, derivatives or debt instruments.
— Brian ArmstrongPrediction markets are more than sports betting
Armstrong says prediction markets are growing extremely fast inside the Coinbase app: within a few months of launch they reached roughly $100 million in annualised revenue, growing more than 100% quarter over quarter. Sports is certainly part of it, and he sees that as entertainment worth celebrating — but he thinks it's much bigger than that, and that the societal implications aren't yet fully understood. His example is policy: to judge whether a policy should be implemented, you could first form a prediction market on "what will the unemployment rate be one, two, three years after implementation". Further out are big questions with no clear settlement date, like "is there a God", which he calls an opinion market, with mechanics like buying a stock — you're just betting on whether it gets better or worse.
— Brian ArmstrongRestore cells to their youthful function
New Limit takes the epigenetic reprogramming route: it doesn't change the cell type (the Yamanaka-lab-style approach does), it only restores the cell's youthful function. The platform first uses AI to search the vast space of transcription factor combinations for candidates, runs large-scale pooled screens, then functional experiments (mostly animal models), and finally human trials. The first batch targets three cell types: hepatocytes, vascular cells and T cells of the immune system. The first Phase 1 trial is in alcoholic liver disease (ALD), where patients have almost no option besides a liver transplant and survival after diagnosis at 12 months is very low. Armstrong estimates that this one drug alone could be worth $20 billion if it succeeds, but the goal is a 40-year-old with the liver, immune system and blood vessels of a 20-year-old.
— Brian ArmstrongWhat America lacks isn't land, it's sandboxes
Armstrong's interest in special economic zones comes from a "meta-question": what is slowing down progress — in many places, it's over-regulation. He mentions that Coinbase invested in Prospera in Honduras, and thinks they've had ups and downs with the local government but are close to resolving it. During the campaign there was discussion of an idea called "freedom cities": the federal government holds vast tracts of land, take 10 or 100 square miles, and on that land you don't have to go through the usual EPA process — you can build data centres, nuclear power plants, unlimited drone delivery without getting FAA approval item by item — first test what works in a safe, controlled sandbox, and once it works, push it back out to the whole country through the normal process. He says the problem now is that in many cases there is simply no sandbox to test in, and if entrepreneurs can't iterate they can't solve the problem.
— Brian ArmstrongIn their own words · checked verbatim
we don't want the AIs to be unbanked you know we want to bank the AIs, they deserve financial services as well
Brian Armstrong5:00
And I believe the stats we last saw on this are that about like 76% of the agent commerce transactions we're seeing are under 30 cents.
Brian Armstrong6:47
That context has to go back into the brain so that you not only a fixix it in this case.But in all future cases going forward.
Brian Armstrong11:42
instead of bugging the team and asking to do something I just send them a PR for a review and it's already done like and that's a magical moment that I just like it's very addictive
Brian Armstrong15:00
People are not being eliminated. I think tasks are being eliminated.
Brian Armstrong17:02
actually, 88% of our revenue is from non Bitcoin trading at this point.So that's an interesting stat, I think.
Brian Armstrong18:33
instead of kind of going after each disease like the symptoms of it we're sort of trying to go for the meta problem of well how do we just restore the function and your cells had when you're younger
Brian Armstrong35:00
You can't really solve a problem as an entrepreneur unless you can iterate on it.
Brian Armstrong44:18
Figures
| Share of agentic commerce transactions under 30 cents | about 76% | 6:47 |
| Minimum per-transaction debit/credit card fee | 30 cents flat plus an additional percentage | 6:47 |
| Share of Coinbase revenue from non-Bitcoin trading | 88% | 18:33 |
| People worldwide without a brokerage or US investment account | about 4 billion | 20:00 |
| Prediction market annualised revenue | about $100 million (within months of launch) | 21:57 |
| Prediction market quarter-over-quarter growth | more than 100% | 21:57 |
| Potential market size for an alcoholic liver disease (ALD) drug | about $20 billion | 31:53 |
Glossary
- Everything Exchange
- Coinbase's goal: equities, commodities, crypto, derivatives and prediction markets all trading in one venue.
- X402
- An agent-to-agent payment protocol incubated by Coinbase and later donated to the Linux Foundation.
- agentic finance
- Letting AI agents hold their own accounts and complete payments and trades themselves.
- epigenetic reprogramming
- Restoring a cell's function to a youthful state without changing its cell type.
How to listen
Founders and investors watching AI agent payments, stablecoins and on-chain assets; engineering leads who want to copy how AI restructures R&D workflows.
The 0:00 cold open is a re-cut clip from the main episode and can be skipped.