Green-Stove Agents

Green-Stove Agents is a team of agents that runs a community solar cooking mini-grid you can walk away from. Off-grid households want to cook on electricity, but a small solar station (3 kW inverter, 6 kWh battery) cannot serve everyone at once. Each household has a different appliance, cooking window and budget. How it works: a Spec Agent turns a one-line goal into a definition of done. Code may tighten it but never loosen it. A Pricing Agent (Nemotron on Crusoe) proposes time-of-day tariffs. Deterministic household acceptance, allocation and battery admission control turn a tariff into bookings. An independent simulator and verifier check every plan against expected and downside solar forecasts. The forecasts come from cited You.com weather data converted to a PV profile. A Supervisor catches looping and drift, proves infeasibility before spending model calls, and stops honestly with PLAN_VERIFIED, NO_VERIFIED_PLAN or INFEASIBLE, plus an operator decision record. Shared Brain (primary track): after each run a Note-writer agent proposes observations. A code Keeper admits only notes backed by recorded evidence. It rejects made-up numbers and rule changes, keeps conflicting notes scoped and expires stale ones. A fresh agent gets a 600-token brief instead of a transcript. Memory is frozen for held-out evaluation, where we compare it against no memory, raw history, last-tariff reuse, nearest-context reuse and deterministic search. Version N+1: a reviewer agent rewrites the Pricing playbook and keeps a change only if it scores better on development days. Honest scope: the grid is a synthetic simulation. Existing Green-Stove hardware is prior work; the agents, memory, verification and evaluation were built today. We ran the full pipeline live on Crusoe (Nemotron 3 Super 120B) with a live You.com forecast. We have not yet measured a memory benefit on live runs, so that is not claimed. Mock results are labeled MOCK.