Tria Lets AI Agents Trade Perpetuals With Restricted Keys

TL;DR

• Tria announced AI-agent access to perpetual trading on October 30.
• Trading keys can expire or be revoked and do not authorize withdrawals.
• The API documentation says both supported venues use real mainnet markets.

Tria has introduced a way for AI agents to open and close perpetual futures positions from a user's account, bringing automated decision-making closer to live trading. Its October 30 announcement says dedicated keys authorize trading without granting withdrawal permission.

The new feature supports Hyperliquid and Decibel. Users can limit how long a key remains active or revoke it. Whether an agent needs approval for a particular trade depends on the connected tool and how it is configured.

Trading authority is already meaningful control

Separating trading from withdrawals addresses one important boundary: an agent does not need permission to send the account's assets to an unrelated wallet merely to place an order. That is a useful distinction in the design of delegated access.

It still leaves the agent with authority capable of changing the owner's financial position. An incorrectly sized trade, repeated order or unintended exposure can lose money without any withdrawal taking place. Keeping custody of an account and controlling every economic decision made within it are different things.

The practical issue is therefore broader than protecting a key. A trading instruction needs an unambiguous market, direction and size, along with a way to determine whether the intended action actually happened. Natural language can make a command easier to express while leaving room for misunderstanding.

The documented connection reaches real markets

Tria's API documentation states that both venues trade on mainnet and that there is no testnet target. It also explains that the available environment setting changes where builder or treasury fees are routed, rather than switching trades into a simulated market.

That is an important distinction for anyone assessing an automated workflow. A configuration that looks like a development setting may still authorize an actual financial transaction. Testing an instruction parser or checking that a tool connects should not be confused with proving a live strategy is safe or profitable.

The documentation also describes access to orders, fills, positions and margin information. Those records can help establish what the system did. They are more useful for accountability than a conversational response that simply says an order was placed.

Stopping new instructions and ending exposure differ

Revoking a key can end an agent's ability to submit further authorized actions. It does not logically follow that every earlier order has vanished or that every existing position has been closed. Access control and the state of the trading account are separate questions.

A well-designed operating process needs to distinguish between stopping automation, cancelling outstanding orders and closing exposure. Each has a different effect. Confusing them can leave a user believing the account is inactive while market risk remains.

Monitoring also needs to cover the whole sequence from instruction to execution. A tool can receive a request, attempt an order and encounter a rejection or partial completion. The economically relevant result is the position and balance that remain, not merely the initial request.

The product test is reliable delegation

AI interfaces may reduce the effort required to interact with trading software, particularly when a task involves several data checks or account actions. That convenience is valuable only if the system preserves the user's intended limits and makes its actions understandable.

Permissions, trade size, approval settings and reviewable records serve different purposes. Restricting one does not automatically supply the others. A trade-only key is therefore one component of a broader control system.

Tria's launch makes that boundary concrete. Users can delegate supported trading actions without handing an agent withdrawal authority, but they are still delegating decisions with real financial consequences. The measure of success will be dependable execution within clear limits, rather than the ability to turn a prompt into a trade.

Written by the News Desk and edited by Samuel Rae.