Coinbase has launched a platform that gives AI agents the ability to do more than just send stablecoins. The new Coinbase for Agents SDK, announced Thursday, lets large language models like ChatGPT and Claude execute trades, manage multi-asset portfolios, and earn yield—all on-chain and without human intervention for each transaction. CEO Brian Armstrong framed it as a natural extension of the company’s existing agent payment tools, but the scope is significantly broader. Where previous integrations focused on enabling agents to pay for API calls or services with USDC, this platform hands agents the keys to a full financial operating system. It is a bet that the agentic economy will not just need a payment rail, but a complete banking and brokerage layer that is natively programmable.
The Architecture: MPC Wallets and Pre-Authorized Spending Limits
The platform’s technical foundation rests on two components that address the core risk of giving software control over money. First, it uses multi-party computation (MPC) wallets, which split private keys across multiple parties so that no single entity—not the user, not Coinbase, not the AI agent—holds a complete key. This means a compromised agent cannot unilaterally drain a wallet. Second, the SDK enforces pre-authorized spending limits and action scopes. A user might grant an agent permission to trade up to $500 per day in a specific set of assets, but not to withdraw funds to an external address. These constraints are embedded at the API level, not as a separate monitoring layer, which means the agent literally cannot exceed them. This is the same pattern we saw in Rain’s Agent Control Layer, but Coinbase is baking it directly into the exchange infrastructure rather than offering it as a third-party overlay.
Beyond Payments: Trading, Yield, and Portfolio Management
The most significant departure from existing agent payment tools is the inclusion of trading and portfolio management. An agent can now rebalance a portfolio across assets, move funds into on-chain yield products, or execute a dollar-cost averaging strategy—all triggered by its own logic rather than a human’s explicit instruction. This transforms the agent from a spending tool into an autonomous financial actor. The implications for treasury management are substantial: a corporate agent could automatically sweep idle USDC into yield-bearing instruments when cash flow projections show a surplus, or hedge currency exposure by executing trades based on real-time market data. Coinbase disclosed that it already processes roughly $1 trillion in annual stablecoin movement and holds $20 billion in on-chain assets, giving it the liquidity depth to support this kind of programmatic activity at scale.
How This Differs from Visa and Mastercard’s Agent Strategies
We wrote last week that Visa and Mastercard are both building a two-layer stack where AI agents use stablecoins for settlement but depend on card-network identity and authorization rails. Coinbase’s approach is fundamentally different. There is no legacy card network in the loop. The agent authenticates via an on-chain wallet, settles in USDC on Base or Ethereum, and operates within permissions enforced by smart contract logic rather than a bank’s compliance department. This is a crypto-native architecture that treats the agent as a first-class economic actor, not as an extension of a human’s card account. The tradeoff is clear: Coinbase sacrifices the regulatory clarity and fraud protection that Visa and Mastercard offer through their bank-partnered models, but gains programmability, lower per-transaction costs, and the ability to support financial operations—like trading and yield farming—that card networks were never designed to handle.
The Open Questions: Liability, Recovery, and Agent Identity
For all its ambition, the platform leaves several hard problems unsolved. The first is liability. If an agent executes a trade based on faulty reasoning or hallucinated market data, who bears the loss? Coinbase’s documentation places responsibility on the user who authorized the agent, but the legal framework for agent-initiated financial errors is untested. The second is recovery. MPC wallets reduce the risk of a single-point key compromise, but they do not solve the problem of an agent making a sequence of individually valid but collectively disastrous decisions. A spending limit prevents a catastrophic drain, but it does not prevent an agent from slowly losing money through bad trades. The third is identity. Unlike Mastercard’s AP4M protocol, which embeds agent identity verification within a permissioned validator set, Coinbase’s model ties agent identity to a wallet address. This is pseudonymous by default, which raises questions about how the system will handle know-your-customer requirements at scale.