71. Mixture of Agents

The lesson

Modern AI systems often need more than one pass of reasoning. A single model may miss an edge case, make an arithmetic error, or choose the wrong interpretation of an ambiguous request. Mixture of Agents, or MoA, is an architecture that improves answer quality by using multiple LLM agents working together in layers.

MoA is not the same as Mixture of Experts, or MoE. MoE happens inside a single neural network: different internal "expert" sub-networks handle different tokens. MoA happens outside the model, across independent agents that each produce a complete response to the same prompt.

The standard MoA pattern has two layers. In the first layer, several "proposer" agents each generate their own answer to the prompt. Because each proposer may use a different model, temperature, or system prompt, the group produces a diverse set of candidate responses. In the second layer, one or more "aggregator" agents receive all the proposer answers and synthesize the best final response.

The aggregator is given explicit instructions, such as: "Review the following candidate answers and produce a final answer that is more accurate, complete, and concise than any individual candidate." This forces the system to self-correct by comparing alternatives, catching errors that a single model might miss.

MoA is especially useful for complex reasoning, coding, mathematics, safety-critical answers, and tasks where no single model is consistently best. It can also be used to combine a cheap fast model for proposals with a strong model for aggregation, trading cost against quality.

The trade-offs are real. MoA increases latency because multiple models must run. It increases token cost because you pay for every proposer and aggregator call. And it adds operational complexity: you must handle failures, timeouts, and disagreements between agents.

Design tips: keep the number of proposers small (3-5 is common), give each proposer a slightly different angle, make the aggregator's criteria explicit, and always evaluate whether the quality gain is worth the extra cost for your use case.

A rail signalling analyst uses TPEE to write an aggregator prompt that reconciles three proposer answers on one incident report, each proposer given a slightly different angle.

A hotel revenue manager gives three proposers different views of a demand forecast and writes one aggregator prompt that merges them into a single recommendation.

A risk analyst at an insurer drafts a mixture-of-agents prompt where three proposers each score a claim summary and one aggregator combines their findings.

An emergency planning officer writes an aggregator instruction that compares three draft response plans and produces one combined plan for review.

A medical researcher uses TPEE to define three proposer prompts on one trial-summary question, plus an aggregator that weighs their evidence before a final draft.

A financial planner drafts a mixture-of-agents prompt with three proposers on a retirement scenario and one aggregator that reconciles their figures.

An audit manager writes an aggregator prompt that reviews three proposer answers on the same control question and sets out where they disagree.

An energy grid engineer designs a small mixture-of-agents prompt, using three proposers on a load-forecast question and one aggregator to combine the answers.

A legal counsel drafts a mixture-of-agents prompt that has three proposers review one clause and one aggregator set out a combined interpretation.

A sports performance analyst uses TPEE to run three proposers on a training-load question and one aggregator that produces a single recommendation for the coach.

Check yourself