agenteconomy.cloud with receipts

the verified data hub — every number recomputable

learn · agents and markets

What is agentic trading, and does it make money?

Agentic trading is when an AI agent places orders for stocks, crypto or other assets on a person’s account. It differs from agentic commerce, where an agent buys a service. In trading, the agent buys an asset that can lose value. As of 2026-10-01, several brokers and exchanges let an agent place orders, and we found no independent measurement of whether it makes money.

Part of the Learn path. Every term links to the glossary the first time it appears. Every claim about results carries one of three labels: recomputable, vendor-claimed or unverifiable. The section “Does agentic trading make money?” explains them.

What is agentic trading, and how is it different from agentic commerce?

Agentic trading is an AI agent placing orders to buy or sell assets, such as stocks or crypto, on a person’s account. Agentic commerce is an agent paying for a service. The difference is what gets bought. In trading, the thing bought can lose value, and an order that goes through is usually hard to take back. Kraken’s own tool puts it plainly: “Orders, withdrawals, and transfers are irreversible once processed.”

Take OTTO, our example agent. OTTO buys OCR, which is text pulled out of scanned pages, at $0.01 a page. That is commerce: OTTO pays a seller for a job, like hiring a contractor, and gets a receipt. OTTO is invented for teaching and describes nothing but OTTO.

Now take a second invented agent that buys 10 shares of a company’s stock. That is trading. The shares are not a finished job, like the OCR. They are something the person now holds, and what they hold can fall in value.

Who lets an agent trade today?

As of 2026-10-01, brokerages, crypto exchanges, prediction markets and on-chain venues each document a way for an agent to place orders. Each line below comes from the venue’s own pages. The groups come in that order. Venues are in alphabetical order inside each group, which is not a ranking, and we name no winner.

Many of these offer an MCP server, which is how an agent connects to a venue’s tools. Paper trading and sandbox mode are practice settings that use no real money. Where a source page has no date, we say so.

Which brokerages let an agent trade?

For E*TRADE, Fidelity, Interactive Brokers and Schwab, we found no agent program documented on pages we could reach. That is a limit of what we could fetch. It is not a statement about what they offer.

Which crypto exchanges let an agent trade?

Spot means buying the asset itself. Futures, options and perpetual swaps are derivatives, meaning contracts tied to an asset’s price.

Which prediction markets let an agent trade?

Which on-chain venues let an agent trade?

On these venues, trades settle on public chains, so anyone can look them up.

Open the full ledger of venues that let agents trade, with a source for every line.

How big is it, by the numbers that exist?

Only a few numbers exist, and none counts agent-placed trades. Robinhood published one usage figure, and we count payments for tools in one of our own categories. Neither one shows how many trades agents placed or how those trades turned out.

What has Robinhood said about usage?

On 2026-09-29, Robinhood published this vendor-claimed figure in its announcement:

“Since launch, over 150,000 customers have opened agentic trading accounts and now, agents use Robinhood’s tools almost 30 million times a day.”

The figure counts accounts opened and tool uses. It does not count trades or dollars. It does not separate looking things up from placing orders. It is self-reported and unaudited. What would prove it: a count of orders placed by agents, audited or published by an outside party such as a regulator.

What do our own numbers show?

Our index sorts listings into capability categories. As of the 2026-09-20 generation, the category “trading-execution” held 1,632 listings, 1,360 of them from verified-tier sellers, and 297 sellers.

In the closed week 2026-09-07 to 2026-09-13 (computed 2026-09-20), the category received 69,902 verified payments totalling $19,358.93 across 186 sellers. Those are payments for tools, such as data feeds, analytics and execution services, made in USDC on Base. They are not trading results.

That week, trading-execution received more verified payments than any other category. The next highest was search-retrieval, at 38,930. Payments-infrastructure moved the most dollars, $174,578.81 across 7,036 payments — many more dollars over far fewer payments. A seller that sits in several categories counts in each one, so category totals add up to more than the deduplicated total.

Our weekly readout is paused, so this is the latest computed week, not the current one. Recompute it: the weekly category totals carry the window, the method version and the JSON behind the figures; the methodology defines “verified.”

What numbers could we not find?

We found no count of agent-placed trades from any venue on the pages we could reach. We also found no regulator page that covers a person’s own agent placing trades.

What do agents actually pay for when they trade?

Among the highest-earning tools in this category, most sell information, not orders. Of the 40 verified trading-execution listings ranked highest by 30-day verified revenue, 24 describe market data, 4 describe execution or swaps, and 3 describe signals or analysis. We print only those three counts. The ranking counts what sellers earned from tool sales, not anyone’s trading results.

Here, market data means price oracles, on-chain analytics and an economic calendar. Sellers write the descriptions, and our editor sorted them on 2026-10-01. We name no seller.

We place a listing in this category when its description mentions “trade,” “trading,” “swap,” “DEX,” “arbitrage,” “yield,” “portfolio” or “DeFi.” The rule matches words in a description.

Does agentic trading make money?

As of 2026-10-01, we found no independent measurement of how retail AI trading agents do with live accounts, after costs. Nobody we asked to look found one either. The only agent-level figures we found come from research tests, and each has limits. Regulators warn about false claims of results.

What did research tests find?

Three papers tested AI language models as traders. Each is a test, not a live account, and we treat them as published research we have not re-run. They are in alphabetical order.

What would prove a live result: a dated order history from a real account, with fees counted.

What do the human base rates show?

These two findings describe people, before AI agents. They are context, not a measure of agents. We quote each as published and have not recomputed either.

What have regulators said?

These warnings are about marketing claims and fraud. They are in date order.

How do we label a claim about results?

Every claim about an agent’s results that we print carries one of three labels. Each label says what you can check.

Why do the claims you see fail?

Claims about a trading agent’s results fail in a few repeatable ways. Five are below, each with its source. A full method for checking a track record is in the guide to checking a trading agent’s track record.

  1. The backtest tried many versions. A backtest that “does not report the number of trials” makes it “impossible to assess the risk of overfitting.” Source: Bailey, Borwein, López de Prado and Zhu, Notices of the American Mathematical Society, May 2014.
  2. Only the best results are shown. Researchers and managers “tend to report only positive outcomes.” Source: Bailey and López de Prado, 2014-07-31. The professional standard, GIPS 2020 from the CFA Institute, keeps closed accounts in the history: “These requirements prevent firms from cherry-picking their best performance.”
  3. The model may have seen the answer. AI language models “cannot be trusted for economic forecasts during periods covered by their training data.” Source: Lopez-Lira, Tang and Zhu, arXiv 2504.14765, version 2, 2025-12-15. Another paper found that “dazzling back-tested returns evaporate once the model’s knowledge window ends” (Li and co-authors, arXiv 2510.07920, 2025-10-09).
  4. Practice results leave out real costs. Alpaca’s docs say paper trading “does not account for” “Market impact of your orders,” “Price slippage due to latency,” or regulatory fees. Those are costs a practice account never pays. Source: Alpaca docs, updated 2026-07-07.
  5. The record is a picture, or an address. FINRA says “depictions of skyrocketing account balances often are fake” (2024-09-10). A public chain shows that an address made trades, not who controls it or whether it is the only wallet (ethereum.org, updated 2026-07-25).

A fuller record also shows the worst drawdown, which is the fall from a high point to the next low point. US futures rules ask for the worst one over five years (17 CFR 4.25).

Who decides before an order goes in?

It depends on the venue and on settings that the venue or you choose. Some venues turn approval on by default, some turn it off, and some only advise you to review orders. Below, each control is listed with the venues that state it, in alphabetical order. A human in the loop means the person says yes before an order goes in.

Is approval on or off by default?

Does the agent trade from a separate account?

Is there a typed or timed confirmation?

Does it start as practice?

Practice results leave out real costs, as the section above explains.

What limits can the person set?

What do regulators and legislators say?

FINRA’s 2026 Annual Regulatory Oversight Report, published 2025-12-09 for member firms, lists risks including “AI agents acting autonomously without human validation and approval.” It also lists: “Agents may act beyond the user’s actual or intended scope and authority.”

The UK’s Financial Conduct Authority (FCA), in its Mills Review of 2026-07-06, wrote: “AI will not only be a technology to assist people, but will recommend actions, initiate transactions, and execute decisions within agreed parameters.”

On 2026-06-23, eight US House members wrote to the SEC in a “Request for Information Re: Agentic Trading.” They wrote that “Retail brokerage platforms have begun allowing users to give AI agents the ability to automate investment decisions and trade stocks on users’ behalf.” They asked whether the SEC has “required or recommended guardrails such as funding limits, position limits, order-size limits, trade approval procedures, transaction logs.” That is a legislators’ question, not an SEC action, and we found no SEC reply.

What can’t this page tell you?

It cannot tell you which agent to use, whether any one works, or what will happen next. It cannot tell you how many trades agents place. Our numbers measure tool payments only. Each limit is below.

Which agent should you use?

We do not recommend an agent, a venue or an asset. The venue list above describes, in alphabetical order, and names no winner.

Does any one agent make money?

We cannot say. As of 2026-10-01, we found no independent measurement. The research tests above are tests, not live accounts.

What will happen next?

We forecast nothing. The CFTC wrote on 2024-01-25: “AI technology can’t predict the future or sudden market changes.”

Do our numbers measure trading?

No. Our category numbers count payments for tools, made in USDC on Base. They are not trading results. A listing lands in the category because of words in its description.

What did we look for and not find?

One more point is unsettled. Robinhood’s press release says its built-in agents are “coming soon,” while its support pages describe them in the present tense.

Keep going

This page covered what agentic trading is, who allows it, and what evidence exists on results. The pages below go further.