A systematic guide to the architectural dimensions that define multi-agent LLM systems. Understand how design choices impact performance, scalability, and coordination.
An autonomous system that can observe its environment, make decisions, and take actions. In multi-agent LLM systems, each agent uses a large language model to reason about what to do. Agents have roles (what they specialize in), objectives (what they're trying to achieve), tools they can use, and memory to store information.
The system that orchestrates multiple agents, manages how they communicate, and controls how they execute tasks. The framework defines the rules: how agents are scheduled, how they share information, and how they coordinate. Different frameworks make different architectural choices, which dramatically affect performance.
Multi-agent LLM frameworks fall into three main paradigms based on how they structure execution and agent interactions. Each paradigm makes different trade-offs between control, flexibility, and performance.
Model your system as a workflow graph. Nodes represent agents or steps, edges define flow. Execution follows the predefined structure step by step.
Best for: Deterministic workflows, predictable execution, transparent debugging.
| Enforces | You Control |
|---|---|
| Workflow structure | Node definitions |
| Edge execution | Memory, data flow |
Organize agents around job descriptions. Each agent has a role that defines what they do. Coordination emerges from agents following their roles and delegating tasks.
Best for: Flexible tasks, adaptive coordination, role specialization.
Examples:
| Enforces | You Control |
|---|---|
| Role abstraction | Role definitions |
| Role interactions | Execution order |
Treat multi-agent systems like simulations. Agents exist in a shared environment and interact with it over time. Coordination happens indirectly through environment state.
Best for: Complex simulations, emergent behavior, environment-mediated interactions.
Examples:
| Enforces | You Control |
|---|---|
| Time-step model | Agent behaviors |
| Environment-mediated | Environment rules |
| Paradigm | Execution model | Coordination style |
|---|---|---|
| Graph-based | Follows predefined workflow graph | Determined at design time |
| Role-based | Role-conditioned reasoning and delegation | Emerges from role interactions |
| GABM | Time-step simulation with environment | Mediated through shared environment state |
These dimensions define multi-agent system architecture. Each dimension represents a fundamental design choice that can be fixed by the framework paradigm or configured by the system designer. We organize them into three categories: how individual agents work, how agents interact, and how the system environment is structured.
These dimensions control how individual agents are designed and how they behave. They determine what agents know, how they think, and what actions they can take.
These dimensions control how agents interact with each other. They determine who can talk to whom, how information flows, and whether agents can collaborate on decisions.
This dimension determines whether system execution is mediated through an explicit shared world state or through agent-local and developer-managed state.