coinbase shipped agent wallets. here's the layer they didn't ship.
Coinbase launched Agentic Wallets in February — agents hold funds, send payments, trade tokens, earn yield, and transact on-chain without waiting for a human to approve each step. it's the right product at the right moment. it's also incomplete in a way that matters for anyone building agents that spend real money.
the problem isn't the wallet. the problem is what surrounds the wallet.
when a human spends money, there's a context stack that's implicit: identity, authorization, spending history, and a dispute path if something goes wrong. when an agent spends money, none of that context exists by default. the agent has a keypair. it can sign transactions. that's it.
Coinbase Agentic Wallets solve the custody problem — the agent can hold and move funds without a human co-signing every call. what they don't solve: how does a merchant know this agent is authorized to spend? what's the agent's transaction history? what happens when an agent hallucinates a payment destination or gets stuck in a retry loop and fires a duplicate settlement?
these aren't edge cases. they're the default state for any agent running in an environment with real financial stakes.
three things that are still missing after you drop in an agentic wallet:
counterparty reputation. a merchant receiving payment from an agent has no way to assess risk before accepting. with a human, you have purchase history, fraud signals, account age. with an agent, you have a wallet address and a JWT. Agent FICO scoring — 300-850, analogous to a credit score but for autonomous agents — gives the receiving side something to evaluate before settlement.
per-session spend controls. "agent holds funds" is too coarse for production use. you want an agent that can spend up to $200 on this task, can only transact with pre-approved counterparties, and auto-expires its authorization after 4 hours. that's not a wallet feature — it's a spend policy layer that has to sit above the wallet.
deterministic settlement confirmation. the current pattern for agent payments is: fire a request, wait for a webhook, retry on timeout. at scale that's a duplicate payment problem. the agent needs synchronous confirmation baked into the transaction response before it proceeds to the next step — not an async webhook it might or might not process correctly.
MnemoPay is built for these three layers: agent-native USDC wallet infrastructure, per-session spending controls, Agent FICO scoring (300-850) for counterparty verification, and settlement confirmation in the response cycle. 672 tests. v1.0.0-beta.1 shipped. 1.4K weekly npm downloads.
Coinbase's infrastructure is the on-ramp. the authorization, reputation, and spend-policy layer is what makes it safe to let agents run without a human watching every transaction.
if you're building agents that spend money, that layer is the difference between a demo and a production deployment.
Top comments (0)