Mixture of Agents

Mixture of agents is an architecture where multiple AI agents — each with different specialisations, models, or configurations — collaborate to handle complex tasks. A routing or orchestration layer analyses incoming requests and directs each sub-task to the most appropriate agent.

How it differs from a single model

A single AI model is a generalist. It handles every type of request with the same architecture, the same training, and the same capabilities. A mixture of agents, by contrast, maintains a team of specialists. One agent might excel at coding, another at creative writing, another at data analysis, and another at research.

Architecture patterns

• Router-based: A central router analyses each request and sends it to the most appropriate agent. The router might itself be an LLM that categorises the request, or a simpler classifier.
• Pipeline-based: Agents are arranged in a sequence where each agent processes the output of the previous one — for example, a research agent gathers information, an analysis agent processes it, and a writing agent drafts the final report.
• Collaborative: Multiple agents work on the same problem independently, and their outputs are synthesised by an aggregator agent. This is similar to ensemble methods in traditional machine learning.
• Hierarchical: A senior agent breaks down complex tasks and delegates sub-tasks to junior agents, then assembles the results.

Why mixture of agents is gaining traction

• Cost optimisation: Route simple queries to cheap, fast models and complex queries to expensive, powerful ones. Most queries do not need the most capable model.
• Quality improvement: A specialist often outperforms a generalist on its specific domain. A coding agent with a code-optimised model produces better code than a general-purpose model.
• Resilience: If one agent fails, others can compensate. The system degrades gracefully rather than failing entirely.
• Scalability: New capabilities can be added by deploying new specialist agents without modifying existing ones.

Real-world examples

• Customer service: Route billing questions to a billing agent, technical questions to a support agent, and sales enquiries to a sales agent.
• Content creation: A research agent gathers information, a drafting agent writes the initial content, and an editing agent polishes the output.
• Code development: Separate agents for planning, coding, testing, and code review.

Challenges

• Routing accuracy: If the router sends a request to the wrong agent, the response quality suffers. Routing is itself a non-trivial AI problem.
• Latency: Multiple agents means multiple inference calls, which can increase response times.
• Complexity: Managing multiple agents, their configurations, and their interactions is significantly more complex than managing a single model.
• Context sharing: Agents need to share context effectively, which requires careful state management.