An AI agent that places orders while you sleep is an automated strategy, and it needs the same checks as any EA or NinjaScript strategy running on a VPS. Robinhood announced Robinhood Agents on 29 September: agents you build in the app that can analyse the market, build strategies and trade from a dedicated agentic account. Trade approval is on by default, and you can turn it off. CoinDesk reports Robinhood's own disclosure that customers assume all risk for trades the agent executes. So the useful question is whether you can explain, afterwards, why it took that one trade.
What was announced
Going by Robinhood's HOOD Summit announcement, there are four parts to it.
- Agents. You build a custom agent inside the Robinhood app, using models from OpenAI and other providers. Robinhood says agents can analyse the market, build strategies and trade on your behalf around the clock.
- Agentic accounts. An agent can only reach the funds in a dedicated agentic account, not the rest of your holdings.
- Trade approval. The default is manual approval, and it can be switched off at any time.
- Loops. A feature that runs a strategy continuously under conditions you set. CoinDesk's coverage describes it as upcoming, and gives the example of an agent checking the market every morning and trading when certain conditions are met.
Note that the in-app agents are "coming soon" for eligible US customers. They aren't something an Australian trader can open today.
CoinDesk also reports the disclosure wording. Customers assume all risk for trades executed by AI agents, and Robinhood "does not control, supervise, monitor, recommend, or audit agents". With Loops, trades may go through without a prompt for each one, including while you're asleep.
That's a plain statement of who is responsible. It's you.
Where this sits next to an EA
An EA in MT5 or a strategy in NinjaTrader is coded rules. You can open the source and read it line by line. If the rule says "go long when the fast EMA crosses above the slow EMA on bar close", then that is what it does, on every bar, every time. Run it twice over the same data in the Strategy Tester and you get the same trades.
That repeatability is what makes testing possible. A coder tests for correct behaviour: here is a bar where it should have traded, did it? Here is a bar where it shouldn't have, did it stay out?
An agent works from an instruction written in ordinary language. The announcement doesn't say how an instruction gets turned into orders, or whether the same instruction on the same market produces the same order twice. So that is the first thing to find out. If it does, you can test it like any other strategy. If it doesn't, you need to know that before it runs unattended, because a rule that can't be reproduced can't be checked against a chart.
What the approval toggle and the separate account cover
Both are sensible controls. They just cover less than they might seem to.
The separate account limits which funds are exposed. Whatever the agent does, it does with the money in the agentic account. That's a cap on the size of a mistake.
The approval toggle limits when an order goes out. With approvals on, the agent can't place an order until you approve it. One click per order is a cheap brake.
Neither one checks that the rule behind the order is the one you meant. An order you approve at 7am because it looks reasonable is still an order you can't trace back to a rule. And once approvals are off, the only thing between the instruction and the market is the instruction.
This is the same position as running an EA you didn't write and can't read. It might be fine. You just can't tell from the equity curve.
Gotchas
No published detail on testing. The announcement doesn't mention backtesting, a simulator or a replay mode for agents. That doesn't mean there isn't one. It means it hasn't been described, so don't assume it.
No published detail on logs. For any automated strategy, the log is what settles arguments: what it saw, what it decided, what it sent, what filled. The announcement doesn't say what record an agent keeps of why it placed an order.
Order types aren't spelled out. Whether an agent's stoploss sits with the broker as a resting order or only exists as part of the agent's instruction matters a great deal when the agent is the thing that has gone wrong.
Loops isn't out yet. Everything above about unattended running applies most to Loops, and the detail may change before release.
What to check before switching approvals off
These are the same checks I'd want on any strategy before it runs on its own.
- Leave approvals on to begin with, and treat every proposed order as a test case. Write down what you expected before you look at what it proposed.
- Check individual trades, not the account balance. For each order: which condition triggered it, and was that condition actually true on the chart at that time?
- Give it the same instruction on different days and see whether the behaviour is consistent. Look especially for the trade it should have taken and didn't.
- Find the record of why. If you can't get a reason for an order after the fact, you can't debug it.
- Keep the agentic account small until the first four are answered.
- Work out where the stoploss lives, and whether it still works if the agent stops.
If you can point to any one trade and say which rule produced it, the agent is doing a strategy's job. If you can't, it isn't ready to run without you.
If you need expert help with an automated trading strategy, reach out to TradingCoders and tell us what you're building.