“AI trading” usually means one of three things: a signal feed you can't interrogate, a copy-trading account you can't audit, or a black-box strategy that asks for your money and your faith at the same time. Agentic trading is a different design entirely — and the difference is worth being precise about.
Agents, not signals
An agentis an AI system with a mandate: a defined job, defined inputs, and a defined set of actions it may take. A signal says “long BTC”. A specialized research agent can say: “given my mandate — funding conditions — here is my thesis, my confidence, and the data I used.” The output is an argument you can read, not a number you must trust.
Agentic trading can organize several specialized agents into a workspace around your intent. In Re:Liquid's shipped product, those agents help you collect and inspect research perspectives before you decide whether to create a ticket.
Why specialized perspectives matter
One model produces one opinion, confidently. Independent perspectives can make it easier to inspect what a single view misses. A useful workspace should make those perspectives visible and give you the context to evaluate them yourself.
- Defined mandates. Each agent has a focused research role, so you can see what question it was asked to investigate.
- Readable evidence. You should be able to inspect the research rather than accept a black-box verdict.
- Human responsibility. The research can inform you, but it cannot replace your judgment or your signature.
The two rules that keep it honest
First: NO-TRADE is always available. Nothing has to happen because research exists. A process that must produce trades will manufacture reasons to trade.
Second: the human decides and signs. Specialized agents can organize research, but they do not take custody, sign an order or place an order on your behalf.
Where the human sits
At the end — holding the pen. In an agentic design done right, AI does not hold your keys and does not execute on its own authority. The workspace organizes the case; you sign the ticket, or you don't. Autonomy applies to research, not to your money.
How to evaluate any agentic product
- Can you read the reasoning, or only the verdict?
- Can you tell what each specialized agent was asked to do?
- Does the product make non-action a valid choice?
- Who holds the keys — and who signs?
- Are research and execution clearly separated?
If a product fails these questions, it isn't agentic — it's a black box with better marketing.