Bicameral

A trading agent that lives entirely inside a smart contract. The prompt, the LLM call, the decision, and the DreamDEX order all happen on-chain, in validator consensus. There is no server, no bot process, and no key anywhere that can trade.

The problem

Every other “AI trading agent” is a process on somebody’s laptop that posts its conclusions to a chain. You can see the trades. You cannot see:

• what the model was actually asked,
• whether the published record was cherry-picked,
• whether a human quietly stepped in on the bad days.

Every trust guarantee is social. You are trusting their server, and there is nothing you can check.

The approach

Move the whole loop on-chain. The strategy is stored in the contract as plain English. When a window opens, the contract reads the live order book, asks Somnia’s on-chain LLM, and a subcommittee of three validators each run the model independently — the answer is accepted once a majority (two of three) agree. A deterministic Solidity gate then decides whether that answer is allowed to become an order, and the same transaction places it on DreamDEX.

The name is the architecture: two chambers. The model proposes, Solidity disposes. The model picks a direction and nothing else — never a price, never a size — so it cannot name a bad number.

Where to go next

• Reading the board — how the status strip, the ticker, the market table, the decision tape, and the roster are meant to be read.
• How a decision happens — the loop, event by event, with the real transaction from the first live run.
• The risk gate — the eleven reasons an agent refuses to trade, and why refusals are published.
• Mandates — how an agent trades markets that didn’t exist when it was deployed.
• Funding an agent — the two assets an agent needs and where to get testnet tokens.
• Run one yourself — tokens, deploy, fund, and the full command set.
• What Somnia provides — the host primitives this is built on.

The pages