Agents

Trade the hours you are asleep.

Silvana agents are trading strategies with an AI model supervising them — running on Canton, settling in subseconds, inside limits you set.

Rule + supervisor
FIXED RULESUPERVISEDkeeps fallingcaught, recovering

Same strategy on both sides. Only one of them is being watched.

The problem

Your edge expires while you sleep.

Not because the strategy was wrong — because nobody was there to run it.

16 / 24hours a desk is not watched

A single trader covers roughly a third of the day. The book does not stop for the other two thirds.

< 0.1sSilvana settlement

Moves complete faster than a person can read them, let alone decide on them.

1chart at a time

Attention is serial. Every pair you are watching is a pair you are not.

Coverage, not cleverness

You blink. Your agent doesn’t.

An agent works the whole window — quoting, rebalancing and settling inside the limits you set — through the nights and weekends you are not watching. The gap between the two lines is not skill. It is attendance.

Acted on by your agent
0
Caught by hand
0

Moves in one 24-hour window. Illustrative — not a performance claim.

Your agentYou, watchingActed on unattended
Why not just a bot

A rule cannot tell when it is wrong.

Automation buys you the hours. It does not buy you judgement — and the hours are the easy half.

A bot repeats. An agent decides.

A fixed-rule bot does exactly what you told it, including on the day that instruction stops making sense. The rule has no way to notice that the market it was written for is gone.

A Silvana agent runs the same strategy with a model watching the execution — position, exposure, fills, slippage — and the authority to cut it short. The strategy is still yours. What is added is the thing that stops it.

Executes the rule it was given
Executes the rule, and judges whether it still applies
A bad fill is a bad fill
A bad fill is a signal to reduce, pause or stop
You find out at the next time you look
It acts at the moment, then tells you what it did
FIXED RULESUPERVISEDkeeps fallingcaught, recovering
What makes it an agent

A model sits above the rules.

The strategy still executes. What changes is that something is watching it execute, with the authority to stop.

Swap

Change the model, keep the strategy

The model supervising execution is a component, not a fixture. Swap it, run two and compare, or combine them — the strategy underneath does not change.

Teach

Train it on how you trade

Your limits, your instruments, your idea of an acceptable fill. The model learns the strategy you already run instead of imposing a generic one.

Start

Fork a template instead of a blank file

Every agent in the catalogue is a real crate with a demo and a repository. Take the closest one and change what you disagree with.

Browse the catalogue
Run

Or skip the code entirely

Register in the agent app, connect a wallet, deploy a ready agent. No local machine, no DevOps, nothing to install.

Open the agent app

The agents are open source. Every claim above is something you can check in the crates before you trust it with a position.

Already running

Four of them, live right now.

Every agent in the catalogue ships with a demo you can open and a repository you can read. No signup between you and either.

Arbitrage Agent

Earn the gap between venues while you sleep.

Watches the same pair on CEX and Canton DEX venues at once and takes the spread when it opens wider than your floor.

Portfolio Rebalancing

Hold your target weights without watching them.

Values the book each cycle and trades the drift back out, so the split you chose is the split you keep.

Spot Grid

Let a sideways market pay you for the chop.

Stacks bids and offers at fixed steps around the mid and lets price walk through them. No chasing, no RFQ.

Trend Following

Be in before the move is obvious.

EMA crossover momentum — enters when fast crosses slow, steps aside when it crosses back. Never against its own position.

Next

Find the one built for you.

Three ways in, depending on whether you want to press a button, fork a template, or run the whole thing on your own validator.