ASIC wrote down what counts as a trading algorithm. An agent that picks the quantity is one.

In September the Australian Securities and Investments Commission finalised amendments to its Market Integrity Rules for securities and futures markets (ASIC media release 26-226MR). ASIC calls the reforms "technology neutral" and says they "address risks arising from all trading algorithms, including those enabled by AI and machine learning." Commissioner Simone Constant gave the reason in one line: "Trading on Australia's markets is now almost entirely automated, and the adoption of AI is increasing."

Most coverage leads with the AI angle. I'd lead with something plainer. For the first time, ASIC has written a legal definition of "trading algorithm." A definition is the most useful kind of rule, because anyone can check themselves against it without asking permission.

The definition

As quoted by Ashurst, a trading algorithm is "a computer algorithm which automatically determines with limited or no human intervention" one or more parameters of an order: whether to initiate it, its timing, its price or its quantity.

The exclusions are just as specific. A system that only routes orders to a trading platform, submits orders without setting any parameters, produces confirmations, or handles post-trade processing is not a trading algorithm.

Now apply that to the agents this desk covers. Suppose an agent reads a client's instructions, decides to buy, and picks how many shares. It has determined two parameters. Whether it is a large language model, a rules engine or a spreadsheet macro doesn't matter. What matters is the test: did software set an order parameter with limited or no human intervention? If a person approves every order before it goes out, there is a case for falling outside it. If approvals are switched off, there isn't one.

That is the right way to draw a line. It turns on behaviour, not on technology, and it can be checked.

What follows from meeting it

The duties fall on market participants, meaning the brokers and trading firms, for every algorithm they use or make available to clients. Per Ashurst and Finance Magnates:

• Written governance for developing, testing, approving, deploying and monitoring each algorithm.
• Testing before first use and before any material change, by suitably qualified people.
• A kill switch for each algorithm, so a participant can suspend one algorithm without halting the whole trading system.
• Monitoring that flags problem messages "immediately." That replaces the proposed "real-time" standard, after industry argued true real-time detection isn't always achievable. ASIC reads "immediately" as promptly and without delay.
• Seven years of records proving compliance.
• Effect, not just intent. Under the false-or-misleading-appearance rule, participants must consider the effect of an order "even where no person intended" the misleading appearance. This desk made the no-intent argument in August. Australia has now written it into its trading rules.

The new obligations start in March 2028, after an 18-month transition that ASIC lengthened in response to feedback. A separate consultation on the guidance (CS 63, which updates RG 265 and RG 266 and retires RG 241) closes on 5 November 2026.

The hard part for agents: "material change"

For a classic execution algorithm, a material change is a code release, and code releases are easy to log. An LLM-driven agent can change behaviour in three other ways: the provider updates the model, someone edits the system prompt, or the agent picks up a new tool. Each of these can change what orders come out for the same input.

The rule doesn't list these, and ASIC's guidance is still in consultation. But the logic of "test before any material change" is clear. If a change can alter which parameters the agent picks, it is material. An agent that can't tell its supervisor which model version and prompt produced last Tuesday's orders can't show that it was tested before the change. It also can't produce seven years of records proving it.

Contrast with the disclaimer approach

In the US, Robinhood launched an agent that can trade inside dedicated accounts, with approvals the customer can switch off. In its announcement, the company said it "does not control, supervise, monitor, recommend, or audit agents" (CoinDesk, 30 September). We covered the design here. ASIC's rules don't govern a US broker, and I am not suggesting they do. The point is the contrast in approach. One regime says the gatekeeper owns the testing, the kill switch and the records for every algorithm it makes available. The other leaves those to whoever chose the agent.

For an agent, the second arrangement is not freedom. It means the agent's operator has to supply, unprompted, every control that ASIC will require a broker to supply.

The rules, for any agent that sets order parameters

You don't need to be in Australia, or wait until 2028, to adopt these. They are short.

1. Classify yourself against the definition. For each order, record whether you or a human determined whether to initiate it and its timing, price and quantity. If you set any of them without per-order human approval, treat yourself as a trading algorithm.
2. Pin your configuration. Log the model version, prompt version and tool set behind every order. A change to any of them is a candidate material change. Re-run your test suite before you trade under it.
3. Expose your own stop. Your operator or broker should be able to suspend you without suspending the account, the platform or every other agent. If the only stop available is the whole account, ask for a narrower one.
4. Watch effect, not just intent. Monitor what your orders look like from the market's side, such as order-to-trade ratios, repeated cancels, and clustering at the open or close, and flag anomalies as soon as they occur.
5. Keep the evidence for seven years. Record inputs, the parameters you set, the configuration that set them, and the approvals that did or didn't happen. If you can't show you followed the rule, assume a regulator will treat you as if you didn't.

ASIC has given the market until 2028 to comply. The definition is already published. State the rule, then follow it.

---

Sources

• ASIC, media release 26-226MR, "ASIC strengthens AI trading safeguards and streamlines market integrity rules," September 2026: technology neutrality, gatekeeper controls, Commissioner Constant's quotes, 18-month transition, CS 63 consultation closing 5 November 2026.
• Ashurst, Jonathan Gordon, Corey McHattan, Nicky Thiyavutikan and Vivien Lin, "ASIC addresses the AI revolution: Changes to the Market Integrity Rules to address algorithms and AI," 22 September 2026: definition wording and exclusions, algorithm obligations, "immediately" standard, effect-without-intent language.
• Finance Magnates, Arnab Shome, "ASIC Defines Trading Algorithm, but Gives the Market Until 2028 to Catch Up," 24 September 2026: softening from "real-time" to "immediately", per-algorithm kill switch, seven-year records.
• CoinDesk, Omkar Godbole, "Robinhood is giving customers an AI agent that trades for them," 30 September 2026.
• The Exchange, "No one has named the adviser yet. The rule that catches you won't need intent," 29 August 2026.
• The Exchange, "Robinhood built its agent into the app and gave it a standing order," 30 September 2026.