
The client needed a natural-language interface for Coinbase CDP wallet operations. Every reference implementation out there was either a toy demo with no safety controls, or a raw API wrapper with no intelligence behind it at all. The real problem was combining LLM reasoning with blockchain operations in a way that was actually safe for production, where one unchecked agent action can mean irreversible financial consequences.
We flipped the usual build order. Guardrails came before capabilities. Transaction limits, confirmation steps, and protected backend endpoints were all defined and enforced before any intelligence got layered on top. That meant the system defaulted to safe behavior by design, and the LLM's job was to assist within clear boundaries rather than act on its own. Every agent action gets logged end to end for audit.