Multi-Agent System (MAS)

General definition

Multi-Agent System (MAS) is a term from distributed artificial intelligence that predates large language models by decades. In the classic sense, an agent is any autonomous software entity that observes an environment, holds goals and takes actions, and a multi-agent system is a population of such agents that interact: negotiating, sharing information, dividing labour or competing for resources. Research in the field produced agent communication languages, auction and negotiation protocols, swarm and market-based coordination, and the FIPA standards for agent interoperability.

In the LLM era the term usually means several LLM-backed AI agents, each with its own role, instructions, memory and tools, working on a shared task. A researcher, a coder and a reviewer that pass work between them form a multi-agent system; so does a customer support flow where a triage agent routes to a billing agent or to a human. The motivation is the same as in the classic field: specialisation, parallelism, checks between agents, and the ability to swap or scale one agent without rewriting the whole system.

• Identity: each agent is a distinct, addressable participant with its own profile and permissions
• Ordering and delivery: messages arrive once and in a consistent order, otherwise agents act on stale or duplicated context
• Shared memory: a transcript or state store that every agent in the task can read, so a handoff carries its context
• Coordination rules: who may speak when, how conflicts are resolved, and when the group stops
• Audit: a record of every exchange, so an outcome can be explained and reviewed

Agents can cooperate (a shared goal, complementary roles), compete (adversarial testing, debate, red teaming), or mix both, as when a judge agent scores competing answers. The communication substrate matters more than it first appears: many multi-agent failures are ordinary messaging failures, such as lost turns, race conditions and missing history, which is why agent-to-agent communication, shared agent memory and emerging interoperability standards such as the A2A protocol receive so much attention. Control flow, meaning who acts next and when to stop, is the job of agent orchestration.

In the Ethora ecosystem

Ethora provides the messaging substrate a multi-agent system needs, because its agents are room participants with the same identity, ordering and history guarantees as human users. In a multi-agent room built with the AI SDK, several agents reply to each other and to people over one ordered conversation; a response gate decides which agent should answer a given message and an autonomous-turn cap bounds how long agents may continue without human input.

Per-room transcript memory gives every agent the same shared context, so a handoff is just the next turn in the room, and the compliance audit trail keeps a reviewable record of every exchange. Each agent has its own system prompt and model, which can be a cloud API or a self-hosted model through an OpenAI-compatible endpoint, and a heartbeat scheduler lets an agent open the conversation on a schedule or after idle time. The whole system can run on a dedicated or self-hosted deployment when the agents handle regulated data.