The two routing layers between an AI agent and your fill price

On May 15, Gemini, the crypto exchange, announced that any AI model speaking Anthropic’s Model Context Protocol (the open standard that lets a language model call external tools as if they were function calls) can now log into a user account, read the order book, and place trades autonomously. The equity side has been creeping in the same direction for weeks. eToro shipped its agent developer store on April 14. Moomoo shipped agent API skills on April 23. Public has been calling itself an “Agentic Brokerage” since the end of March. Schwab is rolling out client-facing AI agents over chat and voice in June. What none of the press releases describes in any detail is how the fill price actually gets set once the agent presses send.

The mechanism, plainly

When a retail order goes through a US broker in 2026, the order has always traveled through one routing decision we do not see. The broker selects which wholesale market maker, often called an internalizer, will fill the trade. The wholesaler pays the broker for the privilege. The broker is then required, under SEC Rule 606, to publish a quarterly report describing where the orders went and how much was paid for them.

In the first quarter of 2026, those payments hit a combined US$1.19 billion across reporting brokers, the largest three-month total since the disclosures began (per the underlying Rule 606 filings, aggregated by industry publications). Citadel Securities alone paid out US$388 million, Robinhood’s two broker-dealer entities together received US$560 million, Charles Schwab received US$355 million, IMC’s Dash venture paid US$227 million, and Virtu Americas paid US$75 million on equities. Per share, the largest payments went to passive non-marketable limit orders at roughly 1.70 cents, against 0.19 cents on plain market orders. The decision about which wholesaler wins the auction for any given order sits inside the broker’s tech stack. We do not pick.

When we add an AI agent in front of the broker, we add a second routing decision before the order ever reaches the broker. The agent now picks which broker API to call, in what order, with what parameters. That choice is made by a language model optimizing for whatever objective sat inside the prompt, against whatever the broker’s API surface happens to expose. Two routing decisions, two sets of incentives.

Why it matters now

Rule 605, the SEC’s execution-quality reporting regime, was amended in 2024 to require large broker-dealers to publish monthly statistics on fill quality starting August 1, 2026. The amended rule introduces a new baseline. Price improvement is measured not just against the headline national best bid and offer, but against the best available displayed price including odd-lot liquidity, and the rule adds a measurement of size improvement. This will be the most detailed public disclosure of retail fill quality the SEC has ever required.

It also audits the broker-to-wholesaler layer only. If an agent calls Schwab on Monday and Interactive Brokers on Tuesday because the Schwab API timed out, that routing choice lives in the agent’s tool-use history and nowhere in any 605 report. If the agent silently downgrades from a passive limit to a marketable limit because a tool returned a partial fill, the broker’s published fill quality is measured on the order the broker received, not on the order the human intended.

Moody’s analysis of 2025 routing data put 27% of US equity volume and 51% of options volume through payment-for-order-flow channels, with wholesale market makers capturing 96% of equity flow and 94% of options flow at the retail layer. Those numbers describe the world that the agent layer is about to plug into.

Two routing decisions, two sets of incentives. Rule 605 audits one of them. The other sits inside a tool-use log.

What we can actually do

The clean test, before delegating a single trade to an agentic broker workflow, is to ask the agent three questions and require an answer for each. First, which broker did it select for this order, and why. Second, what fill spread did it expect, and against which reference price. Third, can it pull the chosen broker’s Rule 606 disclosure for the order type involved and compare the wholesaler’s reported price improvement on that order type against the fill we got. If the agent cannot answer all three, it is optimizing for an objective we never specified.

Closing observation

The thing about hidden routing decisions is that they were already hidden when we placed the order ourselves. Adding an agent does not invent the problem. It moves the problem one layer up. We get back the time we used to spend reading our own statements. We trade it for time we now have to spend reading our agent’s tool-use logs.

The question worth sitting with is whether the August 1 disclosure regime will follow the order chain all the way up to the agent layer, or stop at the broker. That single design choice will decide whether agentic retail trading becomes a transparent improvement or just a more comfortable place to hide the same economics.