Robinhood Lets AI Agents Trade Without Per-Order Approval — and Clients Bear the Losses

At its HOOD Summit in Houston on September 29, Robinhood unveiled Robinhood Agents, allowing AI agents to trade directly on behalf of clients within the app. Under the default setting, every order still requires client approval; but clients can turn off approval and let the agent place orders independently. According to PYMNTS, once an agent makes a bad trade, the loss falls on the client — even if approval has been disabled. The product is aimed at eligible U.S. customers and is "coming soon."

This is not Robinhood's first foray into agentic trading. In May, it launched Agentic Trading, but that solution required clients to bring their own external AI agent and connect it to Robinhood's servers using developer tools. The technical barrier was not low, yet customers still bought in: according to the company's July earnings materials, by the end of July nearly 100,000 customers had opened agent accounts, holding more than $100 million in assets; by the end of September, that number had surpassed 150,000. Unaudited figures released at the summit show that agents call its tools nearly 30 million times per day.

The new product moves these capabilities from the developer interface into the main app, eliminating the need for linking or configuration. The significance of this step lies in distribution: Robinhood had 28.6 million funded accounts as of the end of August, turning agents from a "geek toy" into a default option for the masses.

The approval toggle is both product design and a line of liability

Making "per-order approval" a toggle that can be turned off is equivalent to handing the boundary of authorization to the client to draw themselves. Keep it on, and the human remains part of the decision chain; turn it off, and the agent becomes the de facto executing entity, while the account, funds, and legal liability remain in the client's name. Robinhood's wording is blunt: losses fall on the client.

This is not new in the traditional brokerage framework — clients authorize trades and bear their own profits and losses, which is the basic logic of the brokerage business. But agentic trading pushes this logic to a new position: in the past, "authorization" was a one-time mandate with clear instructions; now authorization is a continuously running software process that can repeatedly place orders during windows when the client is not watching the market, not confirming, or even unaware. The mandate has shifted from a sentence to a piece of code, and who defines the boundaries of that code, who audits it, and whether it can be reconstructed when problems arise — there is currently no industry-wide answer.

For Robinhood, putting agents into the main app is a natural extension of its growth logic: the number of tool calls by agents and the asset scale of agent accounts are both stories worth telling. But the more autonomous the agent, the greater the risk control and compliance pressure on the platform. Once per-order approval is turned off, abnormal trading patterns, cumulative daily order volume, and whether the agent is being manipulated by external prompt injection all need to be identified before execution rather than after the fact. The existing market risk and suitability frameworks at brokerages were designed for the pace of human order placement and may not fit machine behavior that makes 30 million calls a day.

Who is responsible for trades without human confirmation — the industry has no standard yet

From the perspective of Agent payment compliance and security that OceanAlt focuses on, this toggle at Robinhood exposes the securities version of the same problem: when an Agent is granted authority to operate funds, who should perform identity verification (KYA), define the scope of authorization (per-transaction limits, daily cumulative limits, whitelists of tradable instruments), and intercept before settlement — and at what stage?

The payments sector is already moving in this direction — machine payment protocols like x402 write payment intent into the HTTP layer, and stablecoin settlement providers are beginning to attempt screening before disbursement. The securities trading chain is different, but structurally similar: an agent initiates an instruction, the platform executes, funds are transferred, and in between there is a missing "authorization verification" layer independent of both the client and the platform. Robinhood's current answer is to let clients choose: either confirm each order, or bear all consequences. This binary choice can still be covered by disclaimers at the scale of 150,000 accounts and $100 million in assets; if agent accounts continue to penetrate the 28.6 million customer base, regulators are unlikely to be satisfied with the explanation that "the client turned off the switch themselves."

Worth noting is the timeline: external agent access in May, built-in main app in September, and account numbers rising from nearly 100,000 to over 150,000 within four months. Product iteration is clearly outpacing the formation of a liability framework. For companies building agent payment infrastructure, Robinhood's toggle is an observable sample — it proves that demand is real, and it proves that the authorization and attribution layer remains a blank space.

OceanAlt believes that by making "per-order approval" a default-on, one-click-off toggle, Robinhood is essentially substituting a product default for an authorization standard that has yet to take shape: the default determines the actual risk exposure of most clients, while the toggle itself does not generate any auditable authorization record. This means that before regulators provide a unified answer, the authorization boundaries of agentic trading are effectively defined by each brokerage's own product interaction design, and clients are not facing a comparable set of rules, but a set of default settings that cannot be migrated between platforms.

---

Original source: PYMNTS · https://www.pymnts.com/news/investment-tracker/2026/robinhood-lets-ai-agents-trade-without-customer-sign-off/