Beyond Chat: Rethinking Agentic Interfaces for DeFi

This is a placeholder body. The full essay lives at the original on Notion and I am migrating the text into the site. The layout below is here so the post template is real and ready for the words.

We reached for chat because it was the interface we already had. A language model arrives, it talks, so we let it talk to us about our money. That was a convenient default, not a considered choice.

The problem with a blank prompt

A blank prompt is infinite. In a chat box an agent can propose anything, and the person on the other end has to read carefully enough to catch the one action that would drain a wallet. The interface offers no floor and no railing. It asks the user to be the safety system.

Financial actions are not open ended. There is a finite set of things you can do with a position: open it, close it, add margin, take profit, exit. A designed action space names those moves, shows their consequences before they happen, and refuses everything outside the set.

What a designed action space looks like

The agent still reasons in language. It just does not act in language. It acts through a typed surface where every move is legible, reversible where possible, and bounded by limits the user set once and can see at all times.

An agent is only as safe as the world you give it. If the world is a chat box, the world is unbounded.

That is the thesis I keep returning to. More of the argument, with the mechanism worked out, in the full piece.