Hook
JPMorgan, the bank that once called Bitcoin a “fraud,” is now quietly testing AI agents to manage dynamic investment strategies. News broke last week via a Crypto Briefing report: the world’s largest investment bank is deploying autonomous agents that can perceive market signals, reason about risk, and execute trades without human intervention. The reaction from crypto circles has been a mix of awe and fear—awe at the speed of institutional adoption, fear that Wall Street will finally weaponize AI in ways that leave decentralized protocols in the dust.
But I see something else. A deeper betrayal.
Not of crypto, but of trust itself.
Context
JPMorgan has been here before. In 2017, they launched the Quorum blockchain, later sold to ConsenSys. In 2020, their JPM Coin processed billions in wholesale payments. Each time, the narrative was the same: “We are embracing the technology, not the philosophy.” Permissioned ledgers, closed-loop data, private validators. The bank wants the efficiency of distributed systems without the openness.
Now comes the AI agent. According to the analysis I’ve parsed, the system likely combines large language models with reinforcement learning—a hybrid that can ingest messy financial data, form strategies, and self-optimize over time. The technical maturity is early: internal demos, invite-only tests, no connection to public markets. But the direction is clear. JPMorgan is building a self-learning trading brain, housed in their own datacenters, fed by their own proprietary feeds.
From my years working as a decentralized protocol PM—first on 0x relayers in 2017, later on Compound lending models in 2020—I’ve learned one immutable truth: Code is the only permission we truly need. Permissionless systems, where anyone can verify the logic, are the only foundation for trust in a digital world. JPMorgan’s AI agent is the opposite: permissioned, opaque, and unaccountable.
Core: The Architecture of Secrecy
Let me break down what this really means for the financial system.
First, data sovereignty. Every AI agent is only as good as its training data. JPMorgan sits on a mountain of order flow, client holdings, and market-maker privileges. Their agent will learn from that historical data—data that includes the bank’s own previous trading decisions. This creates a feedback loop: the agent will optimize for the patterns that benefited JPMorgan in the past, reinforcing power asymmetries. In DeFi, by contrast, all data is on-chain. Anyone can audit the historical transactions, anyone can build a model on the same open dataset. The playing field is level.
Second, black-box decision-making. The article analysis notes that ‘the specific model is unknown—could be GPT-4 derivative, could be a custom LLM.’ Either way, the reasoning of the agent will be hidden behind corporate walls. When a trade goes wrong—and it will, because markets are chaotic—no one outside the bank will know why. Compare that to a smart contract on Ethereum: every transaction is visible, every liquidation rule is hardcoded and auditable. Trust is not given; it is verified. JPMorgan asks us to trust their engineers, their risk models, their oversight. DeFi asks us to verify the math.
During the Terra/Luna crash in 2022, I isolated myself in the Scottish Highlands. I wrote “The Burden of Belief,” a personal essay about the emotional weight of watching an industry betray its own ideals. That experience taught me that centralization always defaults to secrecy under stress. A centralized AI agent, when faced with a flash crash, will likely halt trading, hide logs, and issue a carefully worded press release. A decentralized protocol would show the chain of failures in real-time, allowing the community to fork or fix. The protocol remembers what the market forgets.
Third, the scalability myth. The analysis estimates that this agent requires moderate compute—perhaps tens of billions of parameters, run on Nvidia H100 clusters. That’s fine for a single bank. But imagine a world where every major bank runs its own secret agent. They’d compete for the same scarce GPU supply, driving up costs and creating a winner-takes-all dynamic. Meanwhile, a DeFi agent can run on any public blockchain, paying gas fees in a competitive market. The aggregated liquidity of Uniswap or Curve is already available to any algorithm willing to pay for it. We build in silence so the network can speak. JPMorgan builds in secrecy so its agents can shout over others.
Contrarian: The Pragmatic Counterargument
Now, let me anticipate the rebuttal. “Ethan,” you might say, “JPMorgan is a regulated bank. They have fiduciary duties. They cannot run their trading on a public blockchain where every order is front-run by MEV bots. Their AI agent is simply a better tool for managing client assets—more efficient, lower costs, better returns. Isn’t that a net positive?”
It’s a fair point. Efficiency matters. DeFi today is still plagued by slippage, latency, and complexity. A centralized AI agent could execute with millisecond precision, avoiding the congestion that plagues Ethereum during peak hours. And yes, JPMorgan’s clients might benefit from improved risk-adjusted returns.
But this argument misses the forest for the trees. The issue is not whether JPMorgan’s agent performs well—it’s whether the system that governs finance should be opaque. The bank is using AI to deepen its moat, not to democratize access. Their agent will be an extension of their existing power: it will trade against retail, against smaller institutions, against you. It will learn from the very data that its parent company collects from millions of customers. That is not liberation; it is surveillance capitalism repackaged as innovation.
Takeaway: The Choice Between Agents
I spent three weeks in 2017 auditing a whitelisting contract for a token sale, only to abandon it and write a 5,000-word essay on why architecture matters more than asset price. I did that because I believed—still believe—that the structure of a financial system determines who benefits from it.
JPMorgan’s AI agent is a structure of control. It will be fast, smart, and profitable—for JPMorgan. But it will not set you free.
The alternative is not a lesser AI. The alternative is an AI agent that runs on open protocols, with verifiable logic, using on-chain data that everyone can see. Imagine an agent that rebalances a lending pool based on transparent interest rate models. Imagine an agent that proposes a new AMM curve and lets the community vote with their liquidity. Liberation is not a promise; it is a state. That state must be built with code that is open, composable, and permissionless.
JPMorgan’s test will likely succeed on its own terms. But the real race is not between a bank and a startup—it is between two visions of trust. One rests on secrecy and authority. The other rests on verification and code. The market may forget which is which, but the protocol remembers.
So let them build their black box. We will build in silence, and let the network speak.