Evergreen: AI Chief of Staff for Startups

Most AI agent systems are vending machines: request, a burst of work, output, then off. Evergreen inverts this. You don't give it a task; you give it an outcome to own, and it owns that outcome for its entire lifetime. The reference build is a Startup Operating System for a fictional SaaS company, Quillo, a form and survey builder. The room is assigned one outcome: keep Quillo healthy and growing. It stays mostly idle, watching. When something real happens, it wakes, judges whether it matters, convenes the right specialists, escalates to the founder only when a decision is needed, records every decision and the reasoning, then goes back to watching. Three pillars make it "Evergreen": persistent (a room lives for weeks; state survives restarts), event-driven (it reacts the moment something happens, not on a prompt), and institutional memory (months later it answers "why did we decide X"). Band is the coordination spine the agents collaborate through during the workflow. A three-tier funnel runs on it: watchers detect (a Metrics watcher derives MRR and catches a drop on its own, not scripted; a Market watcher tracks Quillo's real competitors, Typeform and Jotform), an orchestrator triages materiality and dynamically convenes specialists (Finance, Retention, Competitive Analysis), then stands them down. It is cross-framework by design (a LangGraph orchestrator on gpt-4o, CrewAI specialists on gpt-4o-mini), and every agent's reasoning is powered by the AI/ML API. For example, if the Market watcher detects a pricing change on Typeform, it reports the raw observation to the orchestrator. The orchestrator decides whether it is material; if it is, it convenes the relevant specialists (Competitive Analysis and Finance), which study it in detail and report their findings back to the orchestrator. The orchestrator then sends the founder a single brief with a recommendation. Minor or routine signals are recorded, and the room stays quiet.