Dev.to
7/3/2026

I Run a Team of AI Sub-Agents From a Raspberry Pi. Here's the Architecture.
Short summary
Nova, an AI orchestrator on a Raspberry Pi, manages five specialized sub-agents with scoped tool access to audit, design, and operate systems. The pattern cuts context consumption by 25-35%, enforces safety through infrastructure-level tool gating, and scales complex reviews from sequential to parallel. The post includes concrete decision trees for dispatch timing, production failure modes, and performance tradeoffs.
- •Orchestrated delegation pattern with five role-specific agents (Klaus, Vera, Sasha, Hugo, Yaëlle) each with scoped tools and distinct voices
- •Context window savings of 25-35% at session level by distributing reasoning across fresh contexts, with tool-level safety constraints replacing prompt-level trust
- •Concrete decision tree for when to delegate vs. handle directly, including lessons on over-delegation overhead and parallel task optimization
Generated with AI, which can make mistakes.
Is this a good recommendation for you?



