Web3
Kresus
A conversational trading agent with multi-chain tool execution — built so that no on-chain action happens without an explicit human confirmation first.
The problem
Kresus needed a conversational interface for multi-chain trading — but a chat agent that can execute on-chain actions is also an agent that can move real funds on a misread request. The hard part was never the language model's intelligence; it was authority — deciding exactly what the agent is allowed to do without a human confirming first, and never letting that boundary blur.
The approach
Confirmation is the core feature, not an afterthought bolted on at the end. Every action that touches a wallet — swaps, transfers, contract calls — routes through an explicit confirmation step before execution, with the parsed intent shown back to the user in plain language first. Tool access is scoped per chain and per action type, not granted wholesale to a single do-anything agent.
What we evaluate
Before anything shipped, we built a test set of realistic trading requests, including deliberately ambiguous ones, and checked that the agent asked for clarification or confirmation rather than guessing. That set is re-run on every change to the system, the same regression-gate discipline described in Method.
Status today
Live and in active use for multi-chain trading workflows.
Architecture and evaluation specifics beyond what's above are covered by the client's NDA. Get in touch if you'd like to talk through what's relevant to your own engagement.