Five-Dimensional MAS Collaboration Framework
Overview
A framework for characterizing and designing multi-agent system collaborations across five dimensions. Each dimension represents an independent design choice; changing one doesn’t necessarily change others.
The Five Dimensions
1. Actors: Which agents participate in each collaboration channel? Agents may be homogeneous (same architecture, different prompts) or heterogeneous (different models, capabilities, or tool access).
2. Type: What’s the relationship between agent objectives?
- Cooperation: aligned toward shared goals
- Competition: conflicting individual objectives
- Coopetition: mixed cooperation and competition
3. Structure: How are agents connected?
- Centralized: hub-and-spoke with coordinator
- Decentralized: peer-to-peer direct communication
- Hierarchical: layered with distinct authority levels
4. Strategy: How do agents decide to interact?
- Rule-based: predefined interaction protocols
- Role-based: specialized agents with defined responsibilities
- Model-based: probabilistic reasoning about others
5. Coordination: How static or dynamic is orchestration?
- Static: fixed workflows defined before execution
- Dynamic: adaptive composition based on task state
How to Apply
When designing a multi-agent system:
- Start with the problem: What kind of task? How much uncertainty? How dynamic?
- Choose type based on objectives: If agents have truly shared goals, cooperation. If you need adversarial testing, competition. If negotiation, coopetition.
- Match structure to scale: Centralized for simpler coordination, decentralized for resilience, hierarchical for complex workflows.
- Select strategy for the environment: Rule-based for stable, predictable domains. Role-based for well-understood divisions of labor. Model-based for uncertain, dynamic situations.
- Decide coordination flexibility: Static when the workflow is known. Dynamic when adaptation matters more than predictability.
Limitations
The framework describes what choices exist but doesn’t prescribe which to make, that depends on domain requirements. Some combinations may be underexplored (e.g., competitive + hierarchical + model-based).
The framework also assumes agents can communicate meaningfully. Establishing common ground, shared interfaces, compatible representations, is a prerequisite the framework doesn’t address.
Why This Matters
Most MAS discussions conflate multiple dimensions. “Should we use cooperation or competition?” is a different question from “Should we use centralized or decentralized structure?” Having vocabulary for each dimension enables more precise design conversations and systematic comparison of approaches.
Related:, 05-atom—agentic-ai