Multi-agent systems: from Ferber to executable Conway's Law

Multi-agent systems are not a new topic. Jacques Ferber's founding book was published in France in 1995, under the title Les systèmes multi-agents: vers une intelligence collective. Thirty years later, the industry rediscovers this theme via LLM orchestration frameworks — CrewAI, LangGraph, AutoGen — but often forgets the most important dimension: organization. This article shows how that dimension, theorized by Ferber and Edmund Durfee in the 1990s, becomes production-ready today.

Ferber 1995: an agent does not exist alone

Ferber's central contribution is to rigorously separate two objects: the agent (individual entity, in Wooldridge's sense) and the multi-agent system (the society of agents). To characterize a multi-agent system, Ferber identifies five interdependent dimensions.

• The agents and their properties. Reactive, deliberative, hybrid. With or without mental state. BDI or not.
• The environment. Physical, software, social. Static or dynamic. Accessible or opaque.
• The interactions. Cooperation, negotiation, conflict, competition. Synchronous or asynchronous.
• The organization. Roles, hierarchies, dependencies, power structures. The meta layer that defines who can do what with whom.
• The global dynamics. Emergence of collective behaviour, stability, fault tolerance.

The fourth dimension, the organization, is what distinguishes a multi-agent system from an orchestrator. Without organization, you have agents passing messages. With organization, you have a social structure that constrains and arbitrates these interactions.

Edmund Durfee's contribution: distributed coordination

Parallel to Ferber in France, Edmund Durfee in the United States published on coordination in distributed systems. His contribution is complementary: how, in a multi-agent system, do agents manage to coordinate their actions without a central coordinator? Three mechanisms emerge: social commitment (an agent commits to respect a convention), distributed planning (agents jointly build a plan), and conflict resolution (a protocol arbitrates when two agents want the same resource).

These three mechanisms have been implemented in industrial systems — air-traffic management, industrial process control, military simulation. And they remain the fundamental building blocks of any multi-agent system claiming to go beyond naive orchestration.

The organization test

To assess whether a system is multi-agent in Ferber's sense, you can put five questions to its architecture. If the system cannot answer these questions structurally (and not in the application code of an LLM manager), then it is not a multi-agent system, it is an orchestrator.

• Who decides in case of a conflict between two agents wanting the same resource?
• Who can dismiss an agent that doesn't do its job?
• How does an agent know whom to address for a given task?
• What happens if an agent fails — who takes over?
• How does a new agent join the system and get accepted by the others?

Apply this test to contemporary frameworks. CrewAI: a partial answer via the Manager notion that supervises, but the manager is itself an LLM, therefore fallible. AutoGen: no structural answer, agents converse, the human-in-the-loop arbitrates. LangGraph: an explicit state graph, which lets you answer, but the graph must be coded for each case. None structurally solves the five questions at once.

Conway's Law: the law that makes organization visible

To understand how a system can integrate Ferber's organizational dimension into its very architecture, you need to take a detour through another famous law: Conway's Law, formulated by Melvin Conway in 1968. It states that organizations designing systems inevitably produce designs that copy the organization's communication structure.

This law is generally seen as a fatality — siloed teams produce siloed architectures. But it can also be read in reverse: if you make the organization explicit and executable in the code, you make the social structure of the system inspectable, modifiable, and contestable. That is the idea of Conway's Law executable.

Executable Conway's Law in Swoft

Concretely, in the Swoft platform, the organization of agents is encoded in a routing table: each business domain is associated with an owner agent and a backup agent, with confidence thresholds determining whether the agent can decide alone or must escalate. This table is data, not code, not a prompt. It is modified the way one modifies an HR policy.

Let's check the five-question Ferber test on this implementation. Conflict between agents: the routing table indicates, for each resource, who is owner and who is backup — arbitration is structural. Dismissing an agent: you remove its entry from the table. Finding whom to address: the table serves as a directory. Agent failure: the backup automatically takes over. New entrant: you add an entry and the agent is immediately recognized by the others. The five questions have a structural answer, not an answer encoded in application logic.

Why FAOS and other production NeSy systems fail here

FAOS (Foundation AgenticOS), published in April 2026 on arXiv by Luong Tuan, is one of the most advanced production neurosymbolic systems to date. It deploys 650 agents across 21 industry sectors with asymmetric ontological coupling — role, domain and interaction ontologies constrain agent inputs. It is a significant advance.

But FAOS has no explicit representation of the organizational identity of agents. Agents have ontological roles (admin, user, expert) but no persistent Party identity. They are not in a routing table with confidence thresholds. The system can route a request to a competent agent, but it cannot structurally answer Ferber's five questions.

This absence is exactly the gap identified by Colelough and Regli's 2025 systematic review of neurosymbolism. Among the four gaps in production NeSy systems, they cite multi-agent accountability: when several agents cooperate, traceability of responsibility is generally absent or fragmented. That is precisely what the routing table and dual attribution solve in Swoft.

Conclusion: Ferber made production-ready

Thirty years after the publication of Multi-Agent Systems, the organizational dimension theorized by Ferber is no longer an academic utopia. It becomes production-ready thanks to three combined ingredients: an event-driven architecture that makes inter-agent interactions inspectable, an explicit representation of the organization as data (the routing table), and a shared Party identity that assigns responsibility.

It is this combination that distinguishes a multi-agent system in the strong sense from an LLM orchestrator. And it is what makes it possible to deploy AI agents in regulated sectors where audit, accountability and system stability are non-negotiable.

A multi-agent system without an explicit organization is not a system, it is a herd.

— Restating Jacques Ferber