Skip to main content
Concept technique

Multi-agent coordination

Coordination of AI multi-agent systems

Architectures and patterns for getting multiple AI agents to cooperate without conflict. Three families: monolithic, federated, structurally aligned (Conway).

01 · Qu'est-ce que c'est ?

Multi-agent coordination designates the set of mechanisms by which several AI agents collaborate to accomplish a task neither could complete alone. It is one of the most active research topics in 2024-2026, with rapidly growing literature and multiplying frameworks (LangGraph, CrewAI, AutoGen, Swoft, and many others).

The Cemri et al. paper (arXiv 2025) became a reference by cataloguing fourteen recurring failure modes in multi-agent systems: error cascade, context loss between agents, infinite negotiation, collective hallucination, role contradiction, goal drift, circular-dependency deadlock, etc. Each failure mode has a non-negligible probability of appearing in production without proper architecture.

Three architecture families

  • Star topology: a central orchestrator drives specialized agents. Simple to implement, limited scalability, single point of failure. Example: LangGraph in supervisor mode.
  • Graph topology: peer agents exchanging messages by declared rules. Flexible, hard to debug. Example: AutoGen, CrewAI.
  • Conway-aligned topology: agents structured along the system's organizational or domain boundaries, communication via typed events. More predictable coordination, strong business alignment. Example: Swoft architecture.

02 · Qui est concerné ?

Multi-agent coordination becomes relevant as soon as you go beyond a single agent. In practice, as soon as you address a complete business workflow (subscription, claim, patient file, scoring) requiring several competencies: document extraction, validation, scoring, escalation. For an organization, this is typically the move from a single-agent POC to multi-domain deployment.

03 · Comment Swoft applique ce concept

At Swoft, multi-agent coordination rests on three principles aligned with Conway's Law. First, each agent is attached to a Bounded Context of the DDD metamodel: its scope is an architectural constraint, not a prompt instruction. Second, agents never communicate with each other in free text, but via typed events persisted in the Event Store. Third, long-running workflow orchestration is carried by event-sourced sagas, with automatic compensation in case of partial failure.

This architecture structurally resolves most of the fourteen failure modes catalogued in the literature. Error cascade is bounded by automatic compensation. Context loss is eliminated by the Event Store's shared memory. Role contradiction is impossible because Bounded Contexts are disjoint. Goal drift is caught by the approval gates injected into sagas.

06 · Questions fréquentes

How many agents can be coordinated in practice?
The limit is not in the agent count but in the quality of contracts between them. With a Conway-aligned topology and typed-event communication, you can have dozens of agents collaborating without degradation. With free conversational topology, the system becomes unmanageable beyond five or six agents.
Do you need a central orchestrator?
Not necessarily. An event-sourced saga can play the logical orchestrator role without being a full agent. That is the Swoft approach: the P2P saga carries the orchestration phases (Capture, Classification, Decomposition, Validation, Execution), agents do the work in their respective domains.
How do you audit a multi-agent system?
Three necessary conditions: every agent action is an immutable timestamped event, every event carries dual attribution (who authorized, who executed), and replaying the history reproduces exactly the same result. Without these three, audit stays artisanal and incomplete.
Multi-agent or single enriched agent?
For a simple use case (sales qualification, L1 support), a single agent with access to multiple tools suffices. For a business workflow crossing several organizational responsibilities (subscription, claim, dispensing), multi-agent becomes relevant. Practical rule: one agent per Bounded Context.

Sources officielles

Articles d'analyse

Un projet logiciel Multi-agent coordination ?

Quand Multi-agent coordination demande un logiciel sur-mesure, nous le livrons en quelques semaines, 3× moins cher qu'un éditeur historique.

Décrire votre projet