AI trading agents and your brokerage account

By the TradeAgentic engineering team·Published 30 August 2026·Updated 30 August 2026

Why the brokerage decides what is possible

Most brokerages will let software read an account. Far fewer will let software place an order in one. That single permission is what determines whether autonomous trading is practical for you at all, and it is set by the brokerage rather than by the software you choose.

Without it, the best an agent can do is tell you what it would like to do and wait for you to do it yourself. That is a useful assistant. It is not an autonomous desk.

How software connects to a brokerage

Two ways software connects

The first is conversational. You talk to a model, it reasons about your account, and it can place orders through a connection to the brokerage. It is immediate and it does what you ask.

The second is resident software. It runs continuously against the same account without a conversation, deciding and acting on its own schedule. Nobody types anything. The difference is between asking for a trade and employing something to trade.

What the conversational kind is good at

Investigation. Asking what a position is doing, what the exposure looks like, what happened yesterday. It answers well, and it keeps you in the loop by design, because the loop is a conversation.

Its limit is the same as its strength. It acts when you are present. It has no opinion at three in the morning, it does not notice conditions nobody asked about, and it keeps no record of what it declined, because declining was never on the table.

What resident software is good at

Continuity and discipline. It watches when you do not, applies one standard to every candidate, and records every decision including the ones where it did nothing.

It is worse at answering questions and worse at doing what it is told, because being told what to do is not the point. If you want to intervene constantly, this is the wrong shape of product.

Testing before you commit capital

Paper first, and for longer than feels necessary

Any brokerage worth using offers a paper account that behaves like the real thing. Software worth using should run against one indefinitely, and you should spend that period watching behaviour rather than results.

Watch what it refuses. Watch what it does when data goes stale. Watch what happens when the software is closed with a position open. Those reveal whether it is safe to leave running, which is the question that determines your downside.

What a paper account will not show you

Execution honestly. Simulated fills are kinder than real ones, particularly in thin instruments, so treat paper results as a floor on costs rather than an estimate of them.

Costs that are easy to overlook

Data. Free tiers give a partial view of the market, which is usually adequate for slower strategies and actively misleading for fast ones. Anything trading on short timeframes against incomplete data is deciding on information it does not have.

Fees. Small frequent trades can hand a thin edge entirely to costs. Software worth using accounts for its own costs before deciding a trade is worth taking, rather than discovering it afterwards.

Choosing the right kind

A checklist before connecting anything

Protective orders placed at the brokerage rather than held in the software's memory, because a stop that exists only in a running process disappears when that process does.

Credentials stored in the operating system's own credential store, never in a configuration file, and never transmitted to the software vendor.

Reconciliation against the account at startup rather than trust in its own records, because anything can happen while software is not running and the account is the source of truth.

A hard daily loss limit, and a control that closes everything in one action. If any of those four is missing, stay on paper.

How to tell which kind you are looking at

Read what the product does when nobody is present. If the answer involves you asking it something, it is the conversational kind however it is described. If it has behaviour of its own overnight, it is resident software.

Both get called AI trading agents, and the phrase has stopped distinguishing between them. The behaviour when you are asleep does.

Who each one suits

The conversational kind suits someone who wants leverage on their own decisions and enjoys being in the loop. It is genuinely useful and it is the cheaper answer.

Resident software suits someone who wants positions taken while they are not watching, and who would rather audit a record afterwards than approve anything in advance. That is a different temperament as much as a different product.