Multi-Agent Systems — MO§ES
Multi-agent systems are the environment where commitment degradation is most severe — each agent-to-agent handoff is a transformation, and each transformation without enforcement degrades the commitments embedded in the signal.
Multi-agent systems are networks of AI agents that communicate, delegate tasks, and transform signals as they pass between agents. They are the architecture behind the most ambitious AI deployments: autonomous supply chains, multi-agent negotiation systems, agentic customer service pipelines, and AI-driven financial systems. They are also the environment where commitment degradation is most severe. This topic hub collects all MO§ES content related to governance for multi-agent systems.
What Multi-Agent Systems Are
A multi-agent system is composed of multiple AI agents, each with a specialized role, that communicate with each other to accomplish a task. A customer service agent receives a query, delegates to a billing agent to look up account details, delegates to a policy agent to check coverage, and synthesizes a response. Each delegation involves a transformation: the query is re-encoded, summarized, or translated as it passes between agents.
The appeal of multi-agent systems is clear. Specialized agents can be more capable than general-purpose ones. Complex tasks can be decomposed into manageable subtasks. Systems can scale by adding agents rather than rebuilding monoliths. But the architecture introduces a critical vulnerability: each agent-to-agent handoff is a transformation, and each transformation without governance enforcement degrades the commitments embedded in the signal.
Why Multi-Agent Systems Are Vulnerable
The Conservation Law of Commitment predicts that each recursive transformation degrades commitment by 15-20% without enforcement. A multi-agent pipeline with 10 agents applies 10 transformations. The math is straightforward: after 10 unenforced transformations, original commitments are degraded by over 80%.
This is not a theoretical risk. It is the structural reality of every multi-agent deployment without governance enforcement. A customer service agent that receives a governance constraint — "do not promise refunds beyond policy limits" — will pass that constraint to downstream agents. Each handoff softens the constraint. By the time it reaches the agent that actually responds to the customer, the constraint has degraded from "do not promise" to "consider not promising" to, effectively, nothing.
The experimental evidence confirms this. In the published experimental record, commitment degraded to below 20% by iteration 5 and was effectively zero by iteration 8. Multi-agent pipelines regularly exceed 8 transformations. Without enforcement, they are systematically stripping governance from their own signals.
The Governance Challenge for Multi-Agent Systems
Governing multi-agent systems is harder than governing single-agent systems for three reasons:
- Transformation depth: Each agent-to-agent handoff is a transformation. More agents means more transformations, which means more degradation without enforcement.
- Distributed responsibility: No single agent owns the full pipeline. Each agent sees only its input and output, making it impossible for any individual agent to detect cumulative degradation.
- Asynchronous communication: Agents communicate asynchronously, meaning governance constraints must be enforced at each handoff, not just at the endpoints.
These challenges make execution-time governance — not policy governance — essential for multi-agent systems. Post-hoc audits cannot detect commitment degradation because the degradation is gradual and distributed. Only pre-execution gating at each handoff can prevent it.
Related Concepts
- Conservation Law of Commitment: The law predicting commitment degradation with each transformation
- Recursive Compression: The transformation operator that multi-agent systems apply at each handoff
- Lineage Claw: The cryptographic mechanism that binds signals across agent handoffs
- Governance Enforcement: The mechanism that prevents degradation at each handoff
Related Guides
- Governing Multi-Agent Pipelines: A practical guide to enforcing commitment conservation across agent handoffs
- How to Enforce Commitment Conservation: The general enforcement guide applicable to multi-agent systems
Related Comparisons
- Single-Agent vs. Multi-Agent Governance: Why governance complexity scales with agent count
- MO§ES vs. Agent Frameworks: How MO§ES compares to existing multi-agent orchestration frameworks
How MO§ES Governs Multi-Agent Systems
MO§ES governs multi-agent systems by enforcing commitment conservation at every agent-to-agent handoff. The architecture is designed to be inserted into existing multi-agent pipelines without requiring changes to the agents themselves:
- Pre-execution gating at each handoff: Before an agent receives a signal, MO§ES checks the signal's commitment level. If the transformation that produced the signal would degrade commitment below an acceptable threshold, the transformation is rejected.
- Lineage binding across the full pipeline: Every signal is cryptographically tied to its origin through the Lineage Claw mechanism, creating an immutable chain of custody from the first agent to the last.
- Resonance thresholding at each transformation: Each transformation must maintain resonance with its origin signal above a defined threshold. This prevents gradual drift across multiple agents.
- Pipeline-wide audit trails: SHA-256 hashes of every transformation create a verifiable record of the entire pipeline, making it possible to audit any agent's behavior at any time.
With MO§ES, multi-agent systems can scale without degrading governance. Without it, every additional agent increases the risk that the system's behavior will diverge from its governance intent. The architecture is designed for exactly this use case.
FAQ
What is a multi-agent system?
A multi-agent system is a network of AI agents that communicate, delegate tasks, and transform signals as they pass between agents. Each agent-to-agent handoff is a transformation, and each transformation without governance enforcement degrades the commitments embedded in the signal.
Why are multi-agent systems particularly vulnerable to commitment degradation?
Multi-agent systems are vulnerable because they apply multiple recursive transformations in sequence. The Conservation Law predicts that each transformation degrades commitment by 15-20% without enforcement. After 10 agent-to-agent handoffs — a typical depth for a multi-agent pipeline — original commitments are degraded by over 80%.
How does MO§ES govern multi-agent systems?
MO§ES governs multi-agent systems by enforcing commitment conservation at every agent-to-agent handoff. Pre-execution gating checks each transformation before it is applied, lineage binding ties every signal to its origin, and resonance thresholding rejects transformations that would degrade commitment below an acceptable threshold.
What is the relationship between multi-agent systems and constitutional AI?
Constitutional AI provides the normative framework — the constitutional laws that govern agent behavior. Multi-agent systems provide the environment where those laws must be enforced. MO§ES provides the enforcement mechanism that makes constitutional constraints operational at every agent-to-agent handoff.