ekenx𓅪

Day 3 building Errand for @openservai SERV Hackathon 02 Errand is an agent that researches, books by phone, tops up phones, and buys eSIMs and gift cards, paying sellers through @GoKiteAI Agent Passport and checking every payment before it happens. Ask it to "find a sushi place open Friday at 8 and book a table for two", and it finds the restaurant on Google Maps, calls them with an AI phone agent, and emails you what they said. The part I'm most impressed by is SERV graph sharding. I pasted my payment-guard prompt into the SERV console and it split it into steps: free script steps for the budget maths, Jev decision steps for approve / ask / block, and an agent step that writes the reason in plain words. I call it like a model with one gr_ id, and every run is recorded step by step, so each payment decision can be audited. When v1 was too strict, I could see exactly which step was wrong and fix just that. SERV also plans each errand with a Kronos-audited prompt and reads every paid response with PromptGuard. That's how it caught a Reddit post trying to sneak in a $9 purchase nobody asked for. Kite Passport does the hard part of agent payments well: one passkey tap approves a session with a total and per-payment limit, and the agent can pay any seller in the Kite catalog over x402, even across chains. @reppo closes the loop: every finished run gets an independent grade on how the money was spent, and those grades tell me which guard step to improve next. Still in demo mode: payments simulated, SERV decisions real. Next up: real payments, and Reppo grades comparing guard v1 vs v2.