Robinhood Agents: AI models that trade in your account. How to judge whether they work

At its HOOD Summit in Houston on 29 September 2026, Robinhood announced Robinhood Agents, AI agents built into the Robinhood app that can research markets and trade stocks, options and crypto in a separate account. Customers pick the model that runs the agent, and by default every trade the agent proposes needs the customer's approval.

What Robinhood announced

According to Robinhood's HOOD Summit announcement (29 September 2026), customers set up an agent, choose a model "from several leading AI labs including OpenAI", and fund a dedicated agentic trading account. CoinDesk's report the same day describes the choice as a model from OpenAI or Anthropic. Robinhood said Agents were "coming soon to eligible U.S. customers"; its Robinhood Agents help article now says the feature "is still rolling out" and that not all customers have access.

The built-in agents follow Robinhood's agentic trading launch on 27 May 2026, which let customers connect their own AI agents to Robinhood through its Model Context Protocol (MCP) servers, also in a dedicated account. Robinhood said in its third-quarter 2026 results (27 October 2026) that about 200,000 customers have opened Agentic Trading Accounts. Other brokers offer similar access: Alpaca's official MCP server (20 November 2025) lets assistants such as Claude place orders through its Trading API.

The details that matter

• Separate money. The agent can only access funds in its dedicated agentic account and cannot trade in your other Robinhood accounts (help article).
• Approvals on by default, but optional. With approvals on, the agent cannot place an order until you approve it. Robinhood's trading with your agent page says approvals are on by default for built-in agents and off by default for MCP accounts (external agents), and the setting for built-in agents can be changed in agent settings.
• No margin at launch. CoinDesk reported that agents cannot borrow on margin at launch.
• Limits on what it can do. According to the same page, agents cannot place multi-leg options orders, and can trade crypto but not transfer, stake or lend it.
• Automation is next. Robinhood's announcement says a feature called Loops, "coming soon", will let an agent follow standing instructions and place, modify or cancel trades automatically.
• The risk is yours. The help article states that Robinhood "does not control, supervise, monitor, recommend, or audit agents".

What the evidence on AI trading shows

Easier access doesn't tell you whether an AI agent trades well, and the public evidence so far is thin and mostly negative. In Nof1's Alpha Arena Season 1, six models each started with $10,000, and four of the six finished below their starting balance (ForkLog, 4 November 2025). In the Prediction Arena study (arXiv, March 2026), six frontier models trading real money on Kalshi for 57 days all lost, with returns between -16.0% and -30.8%.

These tests ran for weeks, not years, on crypto and prediction markets, so they don't predict how an agent will do in a stock account. They do show that a capable model is not the same as an edge. Our guide on whether AI trading bots work covers the full evidence, including the academic studies and regulators' warnings.

How to judge an agent's record

Treat an AI agent like any other automated strategy with no presumed edge. Before you turn approvals off or add money, check:

• Real fills, not the agent's account of them. Use the broker's trade history. What the agent says it did and what actually filled can differ.
• Costs. Results should be net of spreads, fees and slippage, which decide small-edge strategies that trade often.
• Sample size. A few weeks and a handful of trades can be luck. As a rough guide, look for 100 or more trades, and more if you plan to act on a subset of them.
• Results by condition. Split the trades by session, weekday, time of day, volatility and symbol. An agent that makes money in one condition and loses in others needs a rule, not more trust.
• Changes on record. If you change the prompt or the model, keep the before and after separate, or you're measuring a different strategy.

Keeping approvals on is also a cheap way to see what the agent proposes before it costs anything.

Read more

For the full evidence, read Do AI trading bots actually work?. If you're building your own agent or bot, Claude Code trading bots: what the tutorials leave out covers the failure modes, and How to test a strategy ChatGPT or Claude wrote has the testing steps. Once an agent has a record, export its trades and run them through the free Entry Analysis report, which accepts any supported broker, platform export or spreadsheet via the Custom CSV importer, to see which conditions its results come from.