Thousands of AI agents on Wall Street. Who actually lets them trade?

Balyasny says it runs thousands of AI agents. JPMorgan expects its agents to work for hours, then days, then weeks. UBS lets agents execute trades. Robinhood customers can remove the per-trade approval altogether.

Those four facts sound like one story. They are not. They describe four very different levels of trust, and the gap between them is where the real decisions sit.

We went through what these firms have said publicly and sorted it into a ladder. The ladder is our framing. The firms did not use it.

Level 1: agents that research

Balyasny Asset Management says its BAMAgent platform supports thousands of autonomous agents working around the clock, building research packages and turning evidence into financial scenarios. In one example, merger-arbitrage work that took three to five days now takes under one day.

The limits are explicit. Agents get only approved tools and data, activity is logged, and output is reviewed before anyone relies on it. In the words of the write-up: "Investment judgment and accountability remain with people." Anthropic case study, September 2026.

Level 2: agents that recommend

Millennium is building a digital risk analyst with Anthropic. It is meant to explain daily changes in risk and form views on exposure, working under the supervision of risk managers, with human experts approving its decisions. It is still being built. Anthropic, August 2026.

JPMorgan plans agents that can run autonomously for an hour or two, screening market activity, client positions and research overnight so private bankers can focus on clients. Its chief analytics officer expects that to stretch to hours, days and then weeks. CNBC, June 2026.

A JPMorgan research note also tested eight agents that shift allocations across market regimes, and all of them beat a 60/40 portfolio on a risk-adjusted basis. Those were backtests, not live trading. InvestmentNews, citing Reuters, July 2026.

Level 3: agents that trade after approval

This is where agents touch the money. At UBS, agents generate thousands of daily alerts for advisers, and once an adviser approves a transaction, they can execute trades and money transfers. The same report says Morgan Stanley keeps human oversight at all times, with agents having no autonomy over portfolio decisions. InvestmentNews, citing Reuters.

Level 4: agents that trade without asking

The top of the ladder is not a hedge fund. It is a retail app.

Robinhood customers can connect an agent to a dedicated trading account. Manual approval is on by default, but it can be turned off, and the agent can then place orders without confirming each one. Robinhood says it does not control, supervise or audit those agents, and that users assume the risk. More than 150,000 customers have opened these accounts. Robinhood, September 2026.

What the ladder shows

This is our reading, and it comes with a limit: we can only see what firms choose to say. They may be doing more in private.

The capability is not the constraint. The same kind of technology appears at every level. What changes is who is allowed to approve what.

Accountability moves as you climb. At Balyasny it stays with investment staff. At UBS an adviser signs off. At Level 4, the customer carries it.

The doors are closing as well as opening. Axios reported in October that major US banks are restricting third-party agents from logging in to customer accounts, with JPMorgan confirming it bars outside parties from using customer credentials. The reporting relies largely on unnamed sources. Yahoo Finance, relaying Axios.

One distinction matters here. Algorithms have traded without a person approving each order for years. The Bank of England's Sarah Breeden drew a similar line in June, separating that from agentic systems that pursue goals with little human supervision. The Next Web. The new question is about software that decides what to do, not software that follows a fixed rule.

Four decisions before you deploy an agent

• What can it read? Name the data and systems, and leave out everything else.
• What can it recommend, and to whom? Decide who receives its output and who is expected to challenge it.
• What can it execute, and who approves? Set this per action, not per agent.
• How do you stop it and explain it? You need a way to halt it, and a record of why it acted.

The interesting question is no longer whether agents can trade. Some already can. It is where you put the gate, and who stands at it.

Where would you put it?

References

• Anthropic: Working at the frontier: How Balyasny Asset Management evaluates and governs Claude Fable 5, September 17, 2026.
• Anthropic: Millennium and Anthropic are building a digital risk analyst with Claude, August 6, 2026.
• CNBC: JPMorgan Chase AI: Bank to deploy more powerful agents this year, June 9, 2026.
• InvestmentNews: Wall Street banks promoting AI agents from research aids into digital coworkers, July 13, 2026, citing Reuters.
• Robinhood: Robinhood Puts the Power of Hedge Funds in Every Trader's Pocket, September 2026.
• Yahoo Finance: Major U.S. Banks Block AI Agents Found in Customer Accounts, October 26, 2026, relaying Axios.
• The Next Web: Bank of England's Breeden warns AI agents could trigger market meltdowns, June 30, 2026.

FinRAG.tech has no connection to the firms named here. The Balyasny and Millennium details come from write-ups published by their AI vendor, Anthropic, and we have not independently verified any firm's claims. This article is commentary, not financial advice.