What is an AI trading agent? Agents, bots and the limits that matter

What an AI trading agent is

An AI trading agent is software that acts on a market on your behalf. It watches prices, decides within the instructions you gave it, and places orders on your account. Unlike a classic trading bot, which you configure through settings, an agent takes instructions in plain language and can explain what it did in plain language too.

The word "agent" matters. It means the software acts for you, not just alongside you. That is exactly why the limits around it matter more than the intelligence inside it.

Why agents arrived in 2026

In 2026 the major exchanges opened their doors to AI agents. Kraken released a command-line tool with MCP support in March and OKX its Agent Trade Kit the same month. Coinbase launched Coinbase for Agents in June, and Binance opened Agent OS in August. Each lets an AI assistant place trades on that exchange.

That made "an AI that trades for you" something any technical user can set up in an afternoon, for free. What none of these tools provides is an agent you can trust with real money: limits it cannot break, a record of why it acted, and honest testing before it starts. Coverage at launch noted that Binance sets no cap on what an agent can lose.

Agent versus bot

• How you instruct it: a classic bot is set up with forms and settings. An AI agent takes plain language.
• What it runs: a classic bot runs a fixed strategy type, such as a grid or DCA. An agent can run any rule you can describe.
• Explanations: a classic bot rarely gives any. An agent can explain each action in your language.
• Main risk: with a bot, the wrong settings. With an agent, acting outside what you meant, without limits.

The best of both is an agent that takes plain language and explains itself, but executes fixed rules you confirmed, inside limits enforced in code.

The risks of letting an AI trade

• No hard limits. An instruction like "do not lose more than $25 today" written in a prompt is a request, not a limit. Nothing stops the model from acting otherwise.
• Variable decisions. Ask a language model the same question twice and you can get two answers. That is fine for drafting and bad for placing orders.
• No standing watch. A general assistant acts when you message it. Between messages, nothing is watching your position.
• No reasoning on record. The exchange sees orders. Unless something records why each order was placed, you cannot audit the agent.
• Key permissions. Connecting an agent with a key that can withdraw puts your whole account at risk.

The limits to insist on

1. A budget the agent can never exceed.
2. A stop-loss on every position.
3. A daily loss limit that pauses the agent when reached.
4. A maximum number of open positions.
5. No leverage, and no averaging down without a cap.
6. A kill switch: one tap to pause, and one to pause and sell open positions.
7. A trade-only key, checked so that withdrawals are impossible.
8. Enforcement in code, before every order, not in the model's instructions.

How TradeCorgi builds an agent

1. You describe the strategy in English, Spanish, Portuguese or Hindi, typed or spoken.
2. The language model drafts a strategy card with the exact rules, and asks only for what is missing.
3. You confirm the card. That version, and only that version, is what runs.
4. It is backtested with real fees and run on paper with live prices.
5. When you go live, a rules engine with no AI in it executes the confirmed rules and checks your limits before every order.
6. The language model explains each trade and each skipped trade in your language.

Editing a strategy creates a new version that starts on paper again; the live version keeps running until you switch.

A worked example

You say: "Buy ETH every Monday with $40, but stop buying if it falls a lot." Your Corgi asks: "If ETH falls a lot, should your Corgi keep buying?" You pick "Pause buys at -15%". The card now says: buy $40 of ETH every Monday at 09:00 UTC, at most $160 a month, pause buys when ETH is 15% below your average price, $20 daily loss limit.

The agent runs on paper, then live on Kraken. One Monday ETH is 17% below your average; the agent skips the buy and writes why. Nothing about that decision depended on the model's mood: the rule said pause, so it paused. This is an illustrative example, not a real result.

Questions to ask any agent provider

• Does the AI place orders, or does it only draft rules that something else executes?
• Where are the limits enforced, and can the agent override them?
• Can the exchange key it asks for withdraw funds?
• What happens when prices are stale, the exchange is down or the key fails?
• Can I see why every order was placed?

If a provider cannot answer these plainly, keep looking.

A checklist before you let any agent trade

• You can read the exact rules it will follow.
• Limits are enforced outside the AI, before every order.
• It has been tested with fees and run on paper.
• The exchange key is trade-only.
• It records a reason for every action.
• You can stop it instantly.