Days after Robinhood, Coinbase gave an AI agent the ability to both move money in a brokerage account and pay for its own data, exposing custody, spend cap, and conflict of interest questions the category will have to answer.
Coinbase's new AI agent can both execute trades and pay for its own premium research, and the structural choice the company made to keep that capability safe is the more interesting part of the launch.
The agent went live on Wednesday, days after Robinhood introduced a comparable capability, and the TechCrunch account of the launch makes clear the build is not a chatbot with a brokerage tab bolted on. Users can hand the agent portfolio rebalancing, autonomous thesis-following, or one-off trade advice. The agent can call Coinbase Advanced, the exchange's pro-trader platform with TradingView charts, to run the analysis. At launch, scope covers crypto spot markets and derivatives. That ceiling will almost certainly move.
The launch becomes a category moment because of the second capability. The agent can spend on external services, including paid research and compute, without a human swiping a credit card. The mechanism is the open x402 payment protocol — built by Coinbase in collaboration with AWS, Anthropic, Circle, and Near, and now governed as an open standard by the x402 Foundation under LF Projects — which lets the agent authorize a payment to a data provider, or to another agent, on its own. There is no login, no subscription, no monthly bill. The agent pays per request.
That is a primitive shift in how agents and money meet. Chat interfaces have always been consequence-free: they suggest, a human acts, and the human's bank is the spending boundary. Coinbase's agent moves the spending boundary into the agent itself, and into whatever wallet funds it. The agent becomes an economic actor, not a conversational layer over a brokerage.
The safety surface is the part worth watching. Coinbase gives users two ways to run the agent: tied to their main Coinbase account, where it can use the balances and instruments the user already trusts the exchange to hold, or inside a separate sandbox, where funds, instruments, and exposure are isolated. The company also plans configurable limits: a maximum trade size, an allow-list of services the agent can call, and a spend cap. Those three controls are the closest thing this category has to a guardrail pattern, and they are arriving before any regulator has asked for one.
The competitive context matters. Robinhood shipped a similar agent — called Agentic Trading — days earlier, with a separate agentic trading account, real-time activity feed, and push notifications for every trade. Two of the largest US retail brokerages, one of which also operates a major crypto venue, putting spending-capable agents in front of consumers in the same week means the question is no longer whether agents will trade and pay on behalf of people, but who brokers the rails they use to do it.
That is where the conflict-of-interest vector sits. Coinbase is the venue where the agent executes trades, the custodian of the wallet that funds the agent, and the original developer of the payments protocol that connects the agent to the research and compute providers it pays. If the agent buys a research report from a Coinbase-affiliated data shop, the exchange is on both sides of that transaction. The sandbox helps. Configurable spend caps help. Neither eliminates the structural position Coinbase now occupies, and the company's incentive to route spend to inside-the-fence providers is now aligned with its revenue model in a way it was not when agents were just chat.
Two things to watch next. The first is whether Coinbase discloses which external research and compute providers the agent is allowed to call, and whether users can audit the payment ledger after the fact. The second is whether the sandbox, as a pattern, becomes the default for this category, or whether competitors ship agents that bind straight to a user's main account and let the user discover the failure modes themselves. The framing Coinbase used to position the launch — that "AI agent traffic now exceeds human internet traffic" — is the kind of claim that needs independent measurement, not adoption, before it carries weight.
For now, the launch is best read as a primitive. A major US exchange has put a spending-capable agent in front of retail users, has split it from the main account by default, and has done so days after a competitor shipped the same idea. The rest of the structure, including who audits the rails, who sets the spend limits, and which research providers get paid, is still being written.