Brian Armstrong (Coinbase): Running a Company With AI Agents and Why Now Is the Golden Age of Building

Coinbase CEO Brian Armstrong on his team of AI agents, Coinbase Brain, fine-tuned open-weight models, agent-to-agent payments, security and the future of finance.

By Tech Team Strategy · Nov 20, 2026 · 7 min read

Brian Armstrong (Coinbase): Running a Company With AI Agents and Why Now Is the Golden Age of Building
Watch the interview

On David Senra's podcast, Brian Armstrong, co-founder and CEO of Coinbase, explains how he rewires his own habits, why he aims for fulfillment instead of happiness, and how he now runs his work through a team of AI agents. He describes Coinbase's internal "Coinbase Brain," why small fine-tuned open-weight models can beat frontier models, why he expects almost all transactions to become agent-to-agent, and why the real prize is the 96% of financial services revenue still held by traditional finance.

Disclosure: the episode includes sponsor messages. We have left those out of this summary.

This article is our own summary, with short quotes, from the video's automatic transcript. Watch the full video above for everything that was said.

The short version

  • Write it down until it sticks. In his 20s he repeatedly wrote that he wanted to start a billion-dollar tech company; within six or seven years of moving to Silicon Valley at about 29, he had. Later, to get past being conflict averse as CEO, he wrote "I love conflict" over and over.
  • Fulfillment over happiness. He calls happiness a fleeting high and fulfillment the deep satisfaction of meaningful work over a long arc. "If it was easy, it'd be boring."
  • He works with seven or eight lead agents and a chief of staff agent; each lead runs three to five sub-agents. He clears their decisions from a prioritized inbox using keyboard shortcuts and voice dictation.
  • Coinbase Brain is an internal system that builds a knowledge base for every employee, team and code repository, refreshed every night. The human job increasingly is to "align the AI" by curating it. Humans still make final decisions and code reviews.
  • Small models can win. A fine-tuned open-weight model that only files and drafts compliance reports outperforms frontier models at that task, and is faster and cheaper.
  • Agents will need money. He expects "99.999%" of transactions to be agent-to-agent within about ten years, running on stablecoin rails, and argues AI must be decentralized the way crypto decentralized finance.
  • Fewer vendors, more security. With cyber incidents roughly doubling year over year, Coinbase is rebuilding some third-party tools in-house to cut risk.
Rate this story
00
On their bookshelfBooks →
  • Zero to One

    Peter Thiel

    David Senra is reading: “Host read this book and found one of his all-time favorite quotes about successful people finding value in unexpected places.”

    Buy on Amazon

People they admire

  • Enzo Ferrari: Admired how passionately he described his products.
  • Steve Jobs: Admired his passionate way of talking about products.
  • David Vélez: Gave him credit for clearly articulating that fintech only captures a tiny fraction of global financial services revenue.
  • Kareem Atiyeh: Values his high standards and 'divine discontent' which helps drive product quality.

As an Amazon Associate, RALECHE earns from qualifying purchases.

Vocabulary23 terms

The words worth stealing from this conversation, what they mean, and who said them.

Elegant words

first principles
Thinking from fundamental truths rather than relying on analogies or established formulas.
David Senra: “They do this by thinking about business from first principles instead of formulas.”
divinely discontent
A state of constant dissatisfaction with current achievements that drives continuous improvement and creation.
David Senra: “David Ogilvy has a great line about describing these people... they're divinely discontent.”
conflict averse
Tending to avoid disagreements, difficult conversations, or confrontations with others.
Brian Armstrong: “And then early on as being a CEO I think I was pretty conflict averse.”
precautionary principle
The policy of avoiding action or restricting technology if its potential consequences are not fully understood.
Brian Armstrong: “So I think the precautionary principle is super dangerous.”
lagging indicator
A metric or sign that changes only after the event or trend it measures has already occurred.
Brian Armstrong: “I think so there's two theories on this one is that it's a lagging indicator.”
red queen race
A situation where one must constantly run and adapt just to maintain their relative position against competitors.
Brian Armstrong: “others are also getting higher productivity that it's like a red queen race like you have to keep running the same faster to stay in the same place.”

AI vocabulary

inference
The process of running live data through a trained AI model to generate a prediction or response.
David Senra: “Router was built to reduce your company's inference costs by matching every request to the most efficient AI model.”
fine-tuning
Training an existing, pre-trained AI model on a specific dataset to optimize it for a particular task.
Brian Armstrong: “I should mention one other thing which is we're been using a lot of openweight models too and actually fine-tuning them.”
frontier models
The most advanced, powerful, and state-of-the-art artificial intelligence models currently available on the market.
Brian Armstrong: “when you're doing the planning, right, you're typically picking like one of the best frontier models that's the most expensive.”
context
The temporary memory space or input data that an AI model references when generating a specific output.
Brian Armstrong: “Then when the agent goes to work on something in that area, it puts the brains into its context.”
openweight models
AI models whose internal parameters are publicly accessible, allowing developers to run and customize them freely.
Brian Armstrong: “I should mention one other thing which is we're been using a lot of openweight models too.”
agentic
Relating to software or AI systems that possess the autonomy to act independently and make decisions to achieve goals.
Brian Armstrong: “especially as we build the economy for agents in gentic finance.”

Tech vocabulary

tokens
The basic units of text processed by an AI model, used to measure data volume and computing costs.
David Senra: “Tokens are money. Save both.”
network effects
A business phenomenon where a product or service becomes more valuable as more people use it.
Brian Armstrong: “Like the liquidity, the network effects from having a like the most liquid exchange in the US.”
B2B
Business-to-business; transactions, products, or software designed to be sold to other companies rather than individual consumers.
Brian Armstrong: “When selling B2B software, you will get a very long list of customization requests.”
air gapping
A security measure that physically isolates a computer or network from any external connections, including the internet.
Brian Armstrong: “That which like air gapping it, distributed, decentralized around the world.”
tech tree
A hierarchical visual representation of possible upgrade paths or technologies a player or organization can unlock.
Brian Armstrong: “there's a tech tree right and you're like I actually have a print out of the civilization tech tree in my office at home”
TAM
Total Addressable Market; the overall revenue opportunity available to a product or service if 100% market share is achieved.
Brian Armstrong: “But also think about the TAM from the agentic economy I just mentioned, right?”

Trends and buzzwords

agentic finance
A financial system where autonomous AI agents manage assets, make decisions, and execute transactions.
Brian Armstrong: “especially as we build the economy for agents in agentic finance.”
humanoid robots
Robots designed to look and move like humans, built to perform physical tasks in human environments.
Brian Armstrong: “I think we need humanoid robots, right? Because it can't move things in the physical world yet.”
agentic economy
An economic system driven by autonomous AI agents transacting and collaborating with minimal human intervention.
Brian Armstrong: “I think that that's going to be like the real-time payment rail for the agentic economy.”
copy trade
A social trading practice where an investor automatically replicates the investment decisions and trades of another person.
David Senra: “They just did this trade. I'll copy trade it.”
Tradfi
Traditional finance; the legacy banking, brokerage, and financial systems operating outside of decentralized cryptocurrency networks.
Brian Armstrong: “but I don't see it happening as much in Tradfi you know.”
Quiz: two questions+10

Get both right to win 10 tokens. One try only.

  1. 1. What percentage of transactions does Brian Armstrong expect to be agent-to-agent within about ten years?

  2. 2. How is Coinbase responding to cyber incidents roughly doubling year over year?

Free account. Your answers are kept while you sign in.

What he said, by topic

Rewiring yourself

Armstrong says repetition isn't "woo woo" manifesting: writing a goal often keeps it top of mind, so you notice opportunities. He compares it to updating the weights of a neural network. "I love conflict" didn't make him seek fights, but he now treats hard calls as meaningful progress and uses anger in a product review as "empowered anger" to raise the bar. People sometimes thank him afterwards for saying what everyone sensed.

Divine discontent

He says he never feels the Coinbase app is good enough and pings 30 or 40 people a day to create urgency. Senra cites David Ogilvy's phrase for such people: "divinely discontent." Armstrong rarely pauses to celebrate: "Cool, that was great. What's next?" He sees fulfillment, not comfort, as the point, and believes everyone should try to contribute to civilizational progress.

Lessons from Civilization

He keeps a printout of the Civilization tech tree in his home office and says he wants to "add a leaf" to it. The game taught him there are many ways to win, and strategies like giving technology to a weaker rival on another continent to keep a future opponent busy. Building a business, he says, scratches the same itch over decades instead of 20 minutes.

How he works with agents

Instead of prompting and getting answers, he gives each agent an area and a goal. An agent might come back with: here are the five biggest opportunities to improve onboarding worldwide; I propose starting with the first three. He clicks yes. Planning uses the best frontier models; sub-tasks go to cheaper ones, and most employees let the system choose the model automatically. He rates today's agents like L5 to L6 employees, available in near-infinite numbers, but says AI is "like a submarine is not exactly like a fish": far beyond humans in some ways, still missing the big picture in others.

The prioritized inbox and a company operating system

Too many open chat sessions overwhelmed him, so Coinbase built an internal queue of decisions, ranked so he can make the most valuable ones in five minutes. He is prototyping an operating system for the company showing priorities, projects, owners, sign-offs and the critical path. In it, AI could flag the biggest risk (an understaffed dependency blocking 14 others) and propose fixes for a leader to approve. The next layer, he says, is VP-level agents: "We need to expand into Brazil. Go do it."

What can't be prompted

Software is getting easier to prompt into existence; regulatory licenses, funded and verified accounts, liquidity and network effects (such as the most liquid US exchange and USDC) are not. Those, he says, are the defensible parts of the business.

Coinbase Brain

The brain ingests A/B tests, user studies, incident reports, meeting transcripts, code reviews and policy documents, distills lessons each night, and is loaded into an agent's context when it works in that area. Humans resolve contradictions ("David said this and Brian said this 14 months ago"). He expects Coinbase, about 4,000 people, to keep roughly the same headcount with each person doing far more.

The agent economy

Agents working across companies will need markets, payment rails and money. Credit cards matter early for compatibility, but he sees stablecoins as the real-time rail for the agentic economy. An economy of countless agents could be far larger than today's human one. He wants Coinbase to be "the most AI enabled financial account," with strict permissions and the customer always in control.

Security and vendors

He says cyber incidents in both crypto and traditional finance have roughly doubled year over year. Coinbase uses frontier agents to review every line of code before production and keeps 99% of crypto offline in distributed, air-gapped storage. Third-party vendors became one of its biggest risks, so it is rebuilding some tools in-house.

Why people fear AI

He traces caution about change to evolutionary psychology, and points to past panics over bicycles, violent video games and civilian cryptography, which Congress almost restricted. He calls the precautionary principle "super dangerous," citing nuclear regulation that halted new plants, and argues for waiting to see real risks before writing rules.

Build tools for yourself or sell them?

Generalizing an internal tool for every company is "10x or 50x" more work, so Coinbase keeps its agent harness internal and expects such tools to become open source. But infrastructure it had to build anyway (connections to every blockchain, custody and staking) is sold through Coinbase Developer Platform, which he compares to AWS. He notes that seven of the world's largest banks are building on it.

The productivity paradox

AI has materially raised productivity at Coinbase but has not yet materially moved revenue. He believes it is a lagging indicator rather than a race where everyone speeds up just to stay in place, partly because traditional finance is adopting it more slowly.

The 96%

Fintech companies together capture about 4% of global financial services revenue; traditional firms hold the other 96%. Inertia keeps people with old habits, like paying rent by paper check, so a product must be dramatically better to make them switch. He thinks AI agents will be that reason, moving up from placing orders, to building positions over time, to buying a basket of assets from a thesis, to following trusted traders: "everybody now has a private banker in their pocket."

What's next

He wants humanoid robots to extend AI into the physical world, from factories and mines to biology labs. At NewLimit, the biotech he co-founded, the test is whether a robot can work as a barista first. And he says his agent "retail app quality" summarized feedback from his post about app problems and assigned five workers to the top issues; a couple of fixes were almost live.

What it means for your business

  1. Give agents goals, not prompts. Assign an area, let them propose options, and decide from a ranked queue.
  2. Build a company knowledge base for your AI. Feed it decisions, incidents and meeting notes so agents work your way.
  3. Try small fine-tuned models for narrow, repetitive tasks. They can be better, faster and cheaper.
  4. Know what AI can't copy. Licenses, trust, liquidity and network effects are your moat.
  5. Cut vendor sprawl. Every third-party tool is a security risk; rebuild where AI makes it cheap.
  6. Expect revenue to lag productivity, and plan how freed-up time turns into growth.

Video details

  • Title: Why Now Is The Golden Age of Building | Brian Armstrong, Coinbase
  • Host: David Senra
  • Guest: Brian Armstrong, co-founder and CEO of Coinbase
  • Published by: David Senra on YouTube
  • Watch: youtube.com/watch?v=NuoEkfIkKX4
  • Note: quotes and figures come from the video's automatic transcript and the guest's own statements, and are lightly cleaned. Check the video for exact wording.

Want to put this to work?

We help owners turn AI trends into a plan, a product and a launch. Book a free call or see our services.

Full transcript

Read the whole interview, word for word, with a timestamp on every answer. Free with a sign-in.

Sign in to read it
Brian ArmstrongCoinbaseDavid SenraAI AgentsCryptoStablecoinsFintechLeadershipFoundersAI

Originally published by Tech Team Strategy

What did you think of this story?

+10 tokens

Tap to vote ↓

GreatAverageNot great

Sign in to vote

Comments

Sign in to commentJoin the conversation and earn 10 tokens.

  • No comments yet. Be the first.

Get the briefing

The week's essential stories, free in your inbox.

More