Robinhood says customers will be able to build an AI agent inside the app and point it at a dedicated trading account. The detail worth your attention is not the agent. It is the default: every trade the agent proposes waits for a human to approve it, and the user can switch that off.
That is a take on autonomy from a very large, regulated front end, and it lands right where onchain agent builders already live.
What Robinhood announced
Per [The Defiant](https://thedefiant.io/converge/cefi/robinhood-unveils-in-app-ai-agents-for-crypto-and-stock-trading), Robinhood is adding in-app AI agents for crypto and stock trading. Trade approvals default to on. A planned feature called Loops would let users run recurring strategies around the clock.
[CoinDesk's coverage](https://www.coindesk.com/markets/2026/09/29/robinhood-adds-ai-agents-perps-and-weekend-trading-in-push-to-win-active-traders) of the announcement at the company's HOOD Summit adds the structure:
- ▸Customers pick a model from OpenAI or Anthropic.
- ▸The agent can analyze markets, build watchlists and trade stocks, options and crypto.
- ▸It works from a dedicated account, and can only access the money in that account.
- ▸At launch, agents cannot borrow on margin.
- ▸Robinhood also points agents at outside data sources, including Unusual Whales, Nasdaq, Token Terminal and Quiver Quantitative.
- ▸Loops is described as standing instructions: monitor markets, then execute a strategy when specified conditions occur. It is still coming, not live.
CoinDesk also reports that more than 150,000 agentic accounts have been created since Robinhood opened its trading infrastructure to outside AI agents earlier this year, with millions of tool calls a day. Those are Robinhood's figures, relayed by the outlet. We have not checked them.
The default is the design
Approve-every-trade is a human-in-the-loop policy, shipped as a product setting. CoinDesk quotes the toggle plainly: users can turn that protection off. A Robinhood director is quoted saying the company expects users "to still have some level of oversight over what their agent is doing."
Read those two lines together. The default says oversight is the starting point. The off switch says the company expects some people to hand over the keys anyway. Both are reasonable, and neither tells you how the agent behaves once approvals are off. Nothing in the coverage says how that mode has been tested, and we are not going to guess.
Builders on the onchain side have been arguing about this exact dial for a year. Spend limits, allowlists, session keys, approval queues: they are all answers to "how much can the agent do before a person looks." Robinhood chose a simple answer for a mass audience: a separate account, no margin, approval on.
The dedicated account is the quiet part
A separate funded account is a blast-radius decision. The agent cannot touch anything outside it, whatever the model decides. If you build onchain agents, you already know the pattern: a purpose-built wallet with a bounded balance beats a main wallet with good intentions.
What the pattern does not give you is a way to check the claim from the outside. On a closed platform, the account boundary is whatever the platform says it is. Onchain, the equivalent is a wallet and a policy you can read. That difference is the whole reason "show me the repo" is still a fair request.
Loops is the part to watch
Approval-by-default works while a person is on the other end of each trade. Loops, as described, is the opposite shape: conditions in, execution out, around the clock. A standing instruction is only as good as its limits, and the coverage does not say what limits Loops will carry. Does the approval default apply to each loop run? Is there a cap per period? Those answers are not in the sources, so they stay open questions.
Anyone building a recurring-strategy agent onchain faces the same list: what triggers it, what bounds it, what stops it, and what record it leaves.
What to watch
- ▸Loops' launch terms. Whether approvals, caps and kill switches apply to recurring runs.
- ▸The approval toggle. How many users turn it off is not something Robinhood has published.
- ▸Model choice. OpenAI and Anthropic are both on the menu, so behaviour will vary by model, and that variance is the user's to manage.
- ▸The onchain mirror. Expect more builders to ship the same pattern with wallet-level controls you can inspect.
If you are building that mirror, start from the stack, not the prompt. Sato maps the wallets, payment rails, skills and MCPs an agent like this draws on, with the Sato Score showing how open, active and checkable each one is. The score is a transparency and liveness signal, not a safety or returns grade. Describe your agent and see what builds it at [satohub.ai/build](https://satohub.ai/build).