The problem JimmyAgent solves

Web3 financial actions are still unsafe and hard to understand for normal users, and the risk becomes even higher when AI agents are allowed to prepare transactions.

Today, users often sign opaque wallet prompts without understanding the real calldata, token approvals, recipients, or contract risks. Browser wallets can be phished, hosted frontends can be compromised, and server-side signing creates dangerous hot-key custody risk. At the same time, AI agents are becoming useful for financial workflows, but giving an agent direct control over private keys is not safe.

Jimmy solves this by separating intent, reasoning, memory, and signing into different trust layers. The user creates intent locally, the AI agent explains and prepares the action, Swarm stores encrypted decentralized memory, ENS provides human-readable identity and agent discovery, and final transaction signing is handled by a satellite-backed SpaceComputer signing layer instead of a normal browser wallet or centralized hot server.

Challenges we ran into

Challenge 1 — Reliable tool-calling for real transactions
Our supervisor uses LLM function calls for ENS checks and Sepolia market actions. The model often put fields like amount, recipient, or to next to the tool call instead of inside payload, which broke our market agent’s expectations. We fixed this in code by merging top-level fields into a single payload before dispatch, and we tightened the system prompt so the model follows a consistent quote → confirm → execute flow. That made swaps, sends, and ENS registration repeatable enough for a live demo.