Five brokers shipped trading agents this week. Do we all end up in the same trade?

In the first days of June, five brokers handed their customers an AI trading agent inside a single week. On June 2, FinTech Global reported that Interactive Brokers had wired Anthropic’s Claude directly into client accounts. Robinhood opened an official connector so an outside agent, whether Claude, ChatGPT, Codex, or Cursor, can place equity orders. eToro, Public.com, and ThinkMarkets shipped their own versions in the same window. The reassuring narrative attached to all of it goes like this: millions of individual investors, each with a different portfolio and a different agent answering a different question, will make markets more diverse, not less. More independent minds, more independent trades, a wiser crowd. I held a softer version of that belief myself. It deserves to be stated in its strongest form, because the strong form is genuinely plausible, and that is exactly what makes it stick.

Why retail flow was supposed to diversify markets

The belief has a real root. Retail flow has long been the diversifying counterweight in a lot of names. A few hundred thousand households deciding to trim, add, or sit tight, each for a small idiosyncratic reason, is close to the textbook picture of uncorrelated demand. When everyone’s reason is different, the orders mostly cancel out and the flow is quiet. Layer an AI agent on top of each of those households and the instinct is that nothing changes about the diversity, because the portfolios are still different, the questions are still personal, and the agent is just a faster way for each of us to act on our own situation. That instinct is half right, and the half that is right is doing all the persuading.

35 million accounts, four shared foundation models

Start with scale. FinanceFeeds put the combined footprint of the five brokers at more than 35 million accounts. Interactive Brokers alone reported roughly 3.9 million accounts and $660 billion in client equity; Robinhood carried about 27 million funded accounts and $250 billion in custody. These are not pilots. The rails are live at a size that matters for flow in mid-cap and small-cap names.

Now look at what is actually shared. Every broker describes the feature as agent-agnostic, which sounds like maximum variety. In practice the agents funnel into a short list of foundation models: Claude, ChatGPT, Grok, Cursor. Four names do most of the reasoning behind tens of millions of accounts. And the brokers do not stop at the model. They publish starter strategies in the launch copy itself: rebalance by sector exposure, screen for companies growing 20% or more a year, run a mean-reversion play that buys oversold names and sells them back on the bounce. Interactive Brokers, operating across more than 170 markets, suggests the same opening questions to everyone, including which trades would move a portfolio toward a target sector weight. The variety we picture lives in the portfolios. The sameness lives one layer up, in the model and the recipe.

Retail already senses this. An Investing.com survey of 938 American investors, fielded in March, found most of them using AI for decisions and a meaningful minority uneasy about where it leads.

The mechanism: the default-prompt monoculture

Here is the part the diversity story misses, step by step. Two investors hold completely different books. One asks the agent to “rebalance by sector,” the other takes the broker’s “screen for 20% growth” starter. Different portfolios, different prompts on the surface. But both requests pass through the same foundation model, which resolves a loose instruction into a concrete list of names using the same priors, the same training data, and the same sense of what “growth” or “overbought” means this week. Feed similar instructions to one model and it returns similar answers, even when the accounts behind them look nothing alike. The convergence does not come from shared holdings or a shared signal. It comes from a shared interpreter and a shared starter recipe, both of which the broker hands out for free at sign-up.

This is a different animal from the convergence that hits quant funds, which comes from shared training data across firms, and from the volatility-targeting herd, which comes from everyone running the same risk arithmetic on the same lookback window. Those are professional plumbing failures. This one is consumer-grade, opt-in, and printed in the marketing.

The variety we picture lives in our portfolios. The sameness lives one layer up, in the model and the recipe the broker handed everyone at sign-up.

To be honest about the limits: no flash event has yet been pinned on retail-agent convergence, and nobody has measured how correlated this flow actually is. The rails are days old. What we can say is that the structure for synchronized retail orders is now live at 35 million-plus accounts, and that the regulator side has already named the worry. IOSCO (the global umbrella body that coordinates the world’s market regulators, whose guidance national watchdogs like the SEC tend to follow), in its May supervisory toolkit, listed herding among market participants and shared service-provider concentration as risks it is monitoring, before the retail rails even shipped.

Pricing the crowding risk on our own accounts

The decision lens here is risk surface, plus a peer-of-mind check on our own behavior. If we run one of these agents, the cheapest protection is to not take the default. The starter strategy the broker suggests is, by construction, the one most likely to be running in the account next to ours. Varying the instruction, the model, and the timing is not about being clever; it is about not being the marginal account that makes a crowded trade one tick more crowded. And when our own statement shows a tidy move into the same names everyone is discussing, it is worth asking whether that was our read or the model’s.