Choosing the Right Multi-Agent Architecture
Key Points
- 1This paper evaluates four multi-agent architectures—sub-agents, handoffs, skills, and router patterns—based on criteria including distributed development, parallelization, multihop conversations, and direct user interaction.
- 2Each architecture presents distinct trade-offs; for example, sub-agents excel in distributed development and parallelization, while handoffs are optimal for multihop conversations and direct user interaction.
- 3The authors advise against multi-agent patterns unless tasks are highly complex, recommending starting with a single agent and selecting an architecture based on specific functional requirements.
The paper, presented by Sydney from Lingchain, discusses criteria for choosing multi-agent architectures, though it initially cautions that a multi-agent pattern might not always be necessary, with complex tasks often being better handled by a single agent with well-designed tools. Multi-agent systems are recommended for increasingly complex problems.
The evaluation framework consists of four core criteria:
- Distributed Development: Assesses the ability of different teams to independently maintain components or agents based on their specialties.
- Parallelization: Measures whether multiple agents can execute concurrently.
- Multihop Conversational Support: Determines if the architecture facilitates sequential calls to multiple sub-agents, carrying context from previous interactions.
- Direct User Interaction: Evaluates if sub-agents can converse directly with the user.
The paper then analyzes four distinct multi-agent architectures:
- Sub-agents (Supervisor Pattern):
- Mechanism: A central supervisor agent orchestrates and coordinates sub-agents, treating them as tools. All message routing is mediated by the main agent, which dictates the invocation sequence and timing.
- Distributed Development: Scores , as it excels in scenarios where different teams manage distinct sub-agents.
- Parallelization: Scores , as agents inherently support parallel tool calling, enabling concurrent invocation of sub-agents.
- Multihop Conversational Support: Scores , being easily organized through iterative cycles of model and tool calling loops.
- Direct User Interaction: Scores low, as direct user interaction with sub-agents is not natively supported or easily achievable, despite technical possibilities via interrupts.
- Handoffs Pattern:
- Mechanism: Agents possess the capability to transfer control to other agents using tool-calling mechanisms. An initial user request goes to an entry-point agent, and subsequent agents can pass control among themselves before generating a final response.
- Distributed Development: Scores low, as it is challenging to develop agents independently when they require specific capabilities to hand off to one another.
- Parallelization: Scores low, as it is not a primary strength of this architecture.
- Multihop Conversational Support: Scores , being exceptionally well-suited for enabling rich, multi-turn conversations with context transfer.
- Direct User Interaction: Scores , noted as potentially the best architecture for allowing direct user interaction with various agents in the chain.
- Skills (Quasi Multi-agent):
- Mechanism: This is presented as a quasi-multi-agent architecture where a single core agent remains in control but dynamically loads specialized prompts and knowledge (skills) as needed. This approach is referred to as progressive disclosure for context management.
- Distributed Development: Scores , as different teams can effectively manage distinct skills based on their expertise.
- Parallelization: Scores . While multiple skills can be loaded and called in parallel, the process is described as two-step, preventing a perfect score.
- Multihop Conversational Support: Scores , as the single core agent can readily make multiple sequential calls to skills.
- Direct User Interaction: Scores , due to the user interacting directly with a single, consistent core agent.
- Router Architecture:
- Mechanism: Involves a routing step that classifies input and directs it to one or more specialized agents. The results from these agents are then synthesized into a combined response. Both the router and synthesizer components can be agentic or deterministic.
- Distributed Development: Scores , as the lack of a standardized protocol (unlike tools or skills) makes distributed development more challenging, though still feasible.
- Parallelization: Scores , as the router can invoke multiple sub-agents concurrently or one at a time.
- Multihop Conversational Support: Scores . This architecture is not designed for multihop conversations as sequential invocations of an agent with state are difficult to manage, suggesting wrapping a stateful router in a tool if needed.
- Direct User Interaction: Scores . The presence of routing and synthesizing layers around agent invocation makes user interaction less direct compared to other architectures.
The paper concludes by emphasizing the importance of starting with a simpler single-agent approach and scaling up to multi-agent patterns only as problem complexity dictates. A summary table consolidating the scores on a five-star scale is provided.