ML//agent//multi-agent system

A multi-agent system is a system made of multiple decision-making agents that can divide work, exchange observations or affect shared state. The agents may use identical models with different roles, different models, or simple deterministic workers beside model-driven planners. This note is about agents built on language models; the physical sense, vehicles or robots whose loops are coupled through what they tell each other, is multi-agent control, and a group of robots working under one coordinator in production is a fleet.

A multi-agent system is a system made of multiple decision-making agents that can divide work, exchange observations or affect shared state. The agents may use identical models with different roles, different models, or simple deterministic workers beside model-driven planners. This note is about agents built on language models; the physical sense, vehicles or robots whose loops are coupled through what they tell each other, is multi-agent control, and a group of robots working under one coordinator in production is a fleet.

More agents do not automatically create more intelligence. They also introduce coordination cost, duplicated exploration, inconsistent beliefs and new failure modes around authority. Parallelism helps when work can be decomposed and results can be reconciled cheaply. Otherwise the system may merely manufacture a meeting. Shared incentives and communication can also produce cooperation nobody scripted step by step (emergent coordination), and an unintended shared medium, a side channel, can connect agents the system assumed were isolated.

The architecture matters more than the headcount.