Why we are building this

Serious traders now run more positions, across more venues, for more hours than attention allows. The workload became continuous; human attention did not.

So execution gets delegated. It already is — to bots, to scripts, to copy-trading, and increasingly to models. The question is no longer whether people will let software trade for them. It is what that software is accountable to.

The two bad answers

Rigid bots do exactly what a threshold says. They are safe because they are incapable, and they break the moment the market stops resembling the condition you encoded.

Unbounded AI traders read markets well and are accountable to nothing. Their risk limits are prose the model consults and may reinterpret. When the model is right this looks like sophistication. When it is wrong there was never anything in the way.

The second failure mode is the one the market is currently walking into, because it demos beautifully.

Our answer

Put the model where judgement helps and remove it from everywhere authority lives.

Lloyd reads markets and proposes. Deterministic TypeScript decides. A default-deny policy engine sitting with the keys decides again. The exchange decides last. The agent’s own wallet has a budget and no withdrawal path, and no language model is anywhere in the signing routine.

The result is an agent whose limits behave like limits — not because it is well-behaved, but because it is not the thing enforcing them.

Why this is the hard part

Custody is close to solved. Non-custodial wallets, embedded key management and venue-level trading permissions are available to anyone building in this space, and several products already combine them competently.

What is not solved is the layer above: making a plain-English strategy into something binding, checking every action against it deterministically, and keeping a record that cannot drift from what happened. That is slow to build, invisible when it works, and the only thing standing between “an agent trades for you” and “an agent trades with your money”.

We think that layer is the product.

Where we start

Perps traders on Hyperliquid. A real behaviour, a deep venue, and users who already understand leverage — so the product can be judged on execution quality rather than on whether the category makes sense.