Coinbase’s Everything Exchange: Agentic Finance, Stablecoins & Tokenization with CEO Brian Armstrong
In a Nutshell
Coinbase CEO Brian Armstrong predicts AI agents will outnumber humans and require their own financial accounts and stablecoin-backed credit cards for micro-transactions under 30 cents, where traditional card fees become impractical. The company is building "everything exchange" infrastructure to trade tokenized stocks, stablecoins, prediction markets, and all asset classes on-chain, while internally deploying AI agents that achieved recursive self-improvement by reviewing and learning from code changes. Armstrong also founded New Limit to pursue epigenetic reprogramming for cellular rejuvenation, targeting a $20 billion alcohol liver disease market initially, with broader applications for restoring youthful function across liver, immune, vascular, and other cell types.
These notes were generated by AI and may contain inaccuracies.
The minimum fee typically on a debit or credit card transaction is about 30 cent fee flat plus the additional percentage. It doesn't really work well for anything under a dollar. It doesn't make sense to pay 30 cents to send a 1 cent payment. About 76% of the Agentic Commerce transactions are under 30 cents.
The AI agents themselves are going to have to have their own accounts. Simple tools exist where with a single command, you just prompt and paste it into your AI agent, and it can now have its own financial account.
There will be more agents than humans in the not too distant future. The agentic economy will be bigger than the human economy at some point in the not too distant future.
Coinbase is at an exciting time in its history because every asset class is coming on chain. It is now possible to trade everything in one place with good liquidity and good cross margin. Stocks, commodities, crypto, derivatives, perpetual futures, and prediction markets can all be traded in one spot in more countries given emerging regulatory clarity.
The everything exchange is the core business. The next big area is stablecoin payments which have been growing even when Bitcoin prices are down. Stablecoin payments are fast and cheap and global, allowing money to be sent at the speed of information in under 1 second for under one cent to move money anywhere in the world.
The third major area is AI changing financial services, called Agentic Finance or AIFI. People with existing financial accounts will want an AI adviser to help build the right portfolio, dollar cost average, tax loss harvesting, and manage investments. The AI adviser can be talked to like a human to help save money on fees and earn yield while cash is sitting.
AI agents are going to need their own accounts. The mission is to bank the AIs so they have financial services. AI agents don't have government ID and can't walk into a bank branch, so self-custodial wallets and crypto rails are used instead of KYC processes.
Both types of accounts are being built. Separate agentic accounts with segregated funds on Coinbase will be possible where agents can be given access to do trading and payments on behalf of users. Self-custodial wallets will also be provided to any AI agent that wants to hold a balance in stablecoin or even raise money and issue a token.
Agents often get stuck at paywalls when trying to spin up AWS resources or book hotels. In the future, agents will need their own spend accounts like having a personal assistant or employee with their own spend account. A stablecoin-backed credit card will be needed for backward compatibility.
X42 is a protocol incubated at Coinbase that was put into the Linux Foundation. Google, Cloudflare, and AWS are working on it. The website edgentic.market shows stats of how many payments are flowing in the new economy and displays all the different services.
Coinbase Business helps companies accept stablecoin payments, which works for both humans and agents. 76% of Agentic Commerce transactions are under 30 cents, making traditional card fees impractical for these small transactions.
A lot of low-value transactions involve gathering information. Venture firms do research compiling from paywalled data. Recruiters scrape data from LinkedIn. Agents talk to other specialized agents to get source data, almost like a tool call.
The edgentic.market website lists top service providers doing search and financial data. Smaller openweight models can be fine-tuned or use reinforcement learning to outperform frontier models on specialized tasks. A compliance review queue trained on 100,000 compliance cases within Coinbase proprietary data allows a small openweight model to outperform a frontier model.
Specialist agents will emerge in the marketplace. One agent might be really good at ordering pizza. Another agent might be an incredible designer trained only on proprietary Figma data.
There will be scarcity for a long time even if the cost of goods and services falls due to robots and AI making commodities, food, and housing cheap. Government printing might create inflationary pressure. Scarcity will remain for things like usable land, energy, and chips. Dyson spheres will be expensive far into the future, so a medium of exchange will still be needed.
Crypto rails are a nice complement for AI to work. Crypto was really good for humans and is going to be essential for AI.
Coinbase started with basics like coding tools hooked into data repositories, machine learning for risk prevention, fraud prevention, and customer support. Recent boundary-pushing work focuses on achieving recursive self-improvement.
Every service, GitHub repository, or backend service can have its own brain. A brain for each team, each service, and each individual can be created, and individuals should be able to take their brain with them when they leave Coinbase.
When an agent makes a change to a service, it ingests the brain for the team and repository. The brain contains the history of all incidents with that service, all financial controls that must be enforced, every AB test ever run, and the full GitHub history of accepted and rejected pull requests.
When a human reviews an agent's change and finds something missed, the correction must go back into the brain so the agent learns for all future cases. This creates a recursive self-improvement system where the accept rate for one-shotted PRs increases over time.
Toshi is the internal agent harness that can make calls out to external vendors and pay using the agentic finance economy over X42 protocols. The CEO asked a high-end expensive model to make a plan for implementing a complex feature broken down into three phases with 10 pieces each. The model was instructed to execute phase one by spinning up 10 separate agents in parallel, recommending the right agent to use including cheaper models.
The model spun up open source models and Grok. Two minutes before the podcast, it pinged that all 10 things from phase one were done and ready for review. Instead of pinging the team in Slack about ideas or bugs, an agent can be tagged in Slack or the request dropped into the internal harness, resulting in a PR for review that is already done.
Productivity is definitely increasing, but this is different from saying average company size will be smaller. Tasks are being eliminated rather than people. Existing companies will not shrink headcount to 10 people. Existing people will be able to get much more done and the pace of everything will accelerate.
Some companies with two or five people will be able to do more than historically thought possible. The velocity of existing teams increases with things shipping faster and more automation. More scalability comes from people, and if cost basis relative to revenue stays constant, there is ability to do dramatically more with the same group, meaning margins should rise over time if revenue grows faster than team size.
88% of Coinbase revenue is from non-Bitcoin trading. Bitcoin is digital gold and sound money that will be around forever, but crypto is updating every aspect of the financial system.
Tokenized stocks were announced and are available outside the US, with work ongoing with the SEC for a US path. This is the first truly tokenized stock product that is not a synthetic or derivative or debt instrument. It is actually a security one-to-one represented by the actual security held in custody.
Tokenization is the broader trend. Stablecoins were the first major use case with a token representing one-to-one a dollar in a bank account or US treasury. Tokenized stocks will give 4 billion people worldwide who lack access to any brokerage or US investment account the ability to invest in companies like Coinbase, Nvidia, and Apple.
Everything will get tokenized including private credit, treasuries, and bank deposits. Financial services will have modern rails. All tokenized assets can be traded on the everything exchange and sent back and forth instantly. A share of stock can be gifted to a nephew by sending it to their wallet without going through complex stock transfer processes. Money can move at the speed of information just like a WhatsApp message.
Tokenization opens up access to bonds, stocks, and other assets that previously had significant friction in purchasing and usage. This development allows global participation in assets that were historically difficult to access. Counter versions of this concept existed in China and other places, where funds were set up for people to buy proportions of stocks listed on exchanges like the Hong Kong Stock Exchange. Global capital flows are expected to change significantly as these tokenization efforts scale. Stablecoins, particularly Tether with USD, represented an early version of this tokenization approach.
Brian Armstrong invested in a company called Harbor seven to eight years ago that was attempting to tokenize real estate, though this was considered too early for the market at the time.
Coinbase launched prediction markets and reached a $100 million revenue run rate within months, with growth exceeding 100% quarter over quarter. While sports betting represents a significant portion of current activity, prediction markets have broader societal implications beyond entertainment.
Prediction markets can inform government policy decisions by allowing bets on outcomes like unemployment rates one, two, or three years after policy implementation. They can also address fundamental questions about human existence, such as debates around evolution versus intelligent design, though these function more as opinion markets without definitive resolution dates.
These markets operate similarly to stock investments, where participants bet on future performance without a specific payout resolution date. The technology enables betting on major life questions and other categories that extend far beyond current applications.
After Coinbase went public in 2021, Brian Armstrong evaluated whether to pursue new ventures while recognizing the company was still at the beginning of its mission to increase economic freedom. Only half of 1% of global GDP currently runs on crypto, with 700 million people having used or held crypto, though monthly active users number between 50-100 million.
Armstrong identified several major technology trends including space, AI, fusion energy, and brain-machine interfaces, but determined these areas already had strong teams working on them. Longevity research stood out as underfunded with fewer quality teams, despite many approaches being viewed as ineffective, such as caloric restriction studies that worked in mice but not humans.
Through a series of dinners with CEOs and scientists, epigenetic reprogramming emerged as a promising area. Shinya Yamanaka's lab demonstrated early results showing cells could be reprogrammed to restore youthful function. The approach focuses on restoring cell function without changing cell type.
New Limit was founded with Jacob Kimmel as CEO, Blake Byers, and Greg Johnson as co-founders. The company has grown to 50-60 people with a lab in South San Francisco. The platform uses AI to test millions of hypotheses about transcription factors for cell reprogramming, building the leading frontier model for epigenetic reprogramming.
The research process involves large pooled screens to identify phenotypic hits, followed by functional assays in animal models, and ultimately human trials. The company has successfully demonstrated reprogramming of at least one human cell type in humanized mouse models and is now testing in non-human primates, with phase one clinical trials launching next year.
New Limit has published its roadmap on newlimit.com, targeting three initial cell types: hepatocytes (liver cells), vascular cells, and immunological cells, specifically T-cells. The approach follows the GLP-1 drug development pattern, starting with high unmet medical need indications.
The first phase one clinical trial will focus on alcohol liver disease, where patients currently have limited options beyond liver transplants and face low 12-month survivability rates. The alcohol liver disease indication alone represents a potential $20 billion market opportunity.
The ultimate goal extends beyond treating specific diseases to developing therapies that restore youthful cell function in healthy individuals. This includes potential applications for liver, immune system, vascular system, brain, muscles, and skin rejuvenation, with skin rejuvenation representing a potential trillion-dollar market.
The strategy targets the root cause of age-related diseases rather than individual symptoms, as most diseases that cause death correlate strongly with aging and loss of cellular regenerative capacity.
Armstrong's primary focus remains on Coinbase's development of agentic finance and stablecoin payments, with the goal of expanding crypto usage from the current 50-100 million monthly active users to over a billion people.
Additional frontier concepts under consideration include cognitive enhancement through various therapeutic approaches targeting adults, children, or embryos. This connects to broader questions about increasing humanity's collective intelligence through both biological and AI-based methods.
Armstrong invested in a company called Preventive that performs embryo editing for disease prevention. Research indicates 80% of Americans support embryo editing for disease prevention, suggesting this technology may become normalized over time, similar to how seat belt usage became standard.
The discussion also addressed concerns about whether genetic enhancement packages might reduce neurodiversity, with Armstrong arguing that unlimited diversity in human preferences would likely increase rather than decrease diversity in outcomes.
Special economic zones represent meta problems that could unlock progress across entire categories of innovation. Overregulation slows down progress in many parts of the world, with Europe serving as an example, while the US has its own regulatory challenges. Various solutions have been discussed including seasteading, Mars colonization, and brain upload into cyberspace.
Coinbase made an investment in Prosper, a special economic zone project in Honduras, to learn more about this approach. The project has experienced mixed results with the local government but appears to be approaching a resolution. Interest in special economic zones was higher before the most recent Trump administration because the deregulatory environment has created significant opportunities within the US.
An idea discussed during the Trump campaign involved creating "freedom cities" on federal land. The federal government owns various pockets of land across the country, and the proposal suggests designating 10 to 100 square miles for experimental zones where normal regulatory processes would be streamlined.
These zones would allow experimentation with technologies like nuclear power, drug discovery, crypto, and drone delivery without requiring standard approvals from agencies like the EPA or FAA. The concept aims to create safe, contained environments where innovation can be tested and iterated upon, with successful experiments potentially being brought to the rest of the country through normal regulatory processes.
The current problem is that many innovations lack a sandbox environment for testing, which stifles entrepreneurial problem-solving since iteration is essential for solving complex challenges.
Successful special economic zones globally include Shenzhen, Singapore, and Dubai, which have driven significant economic growth. The US already functions as 50 special economic zones due to federalism and the state system, with this competitive structure having allowed the country to maintain global leadership. However, there's potential to expand this approach further with additional special economic zones.
Brian Armstrong expressed enthusiasm about implementing special economic zones or "pieced cities" in the US to accelerate innovation and economic growth.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
