Mixture-of-Agents (MoA) is an architecture for getting a better answer out of several large language models than any one of them produces alone. It was introduced by Wang et al. in 2024 ("Mixture-of-Agents Enhances Large Language Model Capabilities"), and the core idea fits in a sentence: have multiple models draft answers, let them read each other's drafts, and have a final model synthesize the result.
That sentence hides three distinct roles, so it's worth pulling them apart.
The three parts: proposers, layers, aggregator
Proposers are the models that generate candidate answers. Each proposer receives the same prompt and answers independently, in parallel. The proposers are deliberately heterogeneous — different model families, trained on different data, with different failure modes. That diversity is the fuel for everything that follows.
Layersare rounds of refinement. In the first layer, proposers only see the user's prompt. In every subsequent layer, each model sees the prompt plus all of the previous layer's drafts, and produces a revised answer informed by them. This is the step people miss when they summarize MoA as "ask several models and pick one." The models aren't just voting — they are reading one another's work and reacting to it.
The aggregatoris the final model in the stack. It receives the prompt and the last layer's refined drafts, and writes the single answer the user actually gets. It isn't choosing a winner; it's synthesizing — keeping the strongest reasoning from one draft, a correction from another, discarding claims that only one model made and none could defend.
Why cross-critique beats simple ensembling
Classical ensembling — run N models, take the majority answer — works when outputs are discrete, like a classification label. Text generation breaks it: four models produce four different paragraphs, and there is no "majority paragraph" to vote for. You could pick the draft a reranker scores highest, but then your ceiling is the best individual draft.
MoA's refinement layers raise that ceiling. The paper's authors observed what they call the collaborativeness of LLMs: models produce measurably better answers when shown other models' attempts, even when those attempts are worse than what the model would produce alone. A weak draft can still contribute a fact, an angle, or a counterexample the stronger model missed. And errors tend to be uncorrelated across model families — a hallucinated citation from one proposer rarely survives a layer in which three other models, trained on different data, get to read it. Majority voting can only select among existing answers; cross-critique produces an answer that didn't exist before.
The headline result
The striking number from the paper: an MoA configuration built entirely from open-source modelsscored 65.1% on AlpacaEval 2.0, ahead of GPT-4 Omni's 57.5% — at a time when GPT-4 Omni was the strongest single model on that leaderboard. None of the panel's members could beat GPT-4 Omni on their own. The mixture did. Similar gains showed up on MT-Bench and FLASK. The result has the usual benchmark caveats — AlpacaEval measures instruction-following preference, not everything you care about — but the direction is the point: orchestration added capability that no individual member had.
The honest caveats
MoA is not free, and it is not always the right call.
- Latency. Each layer must finish before the next begins. Calls within a layer run in parallel, so a propose–critique–aggregate run costs roughly the wall-clock time of two to three sequential model calls, not one. For a chat UI where users expect the first token in under a second, that can be disqualifying.
- Cost. You are paying for every draft and every critique, not just the final answer. A four-proposer, two-layer run consumes roughly an order of magnitude more tokens than a single call. (We work through the exact arithmetic in a separate post.)
- Diminishing returns on easy prompts.If the task is "extract the invoice date," four frontier models will produce four identical drafts and you've paid 4x for confirmation, not improvement. MoA earns its cost on open-ended reasoning, drafting, and review — tasks where models genuinely disagree.
A reasonable rule: use a mixture where you'd want a second opinion from a human colleague. Use a single model where you wouldn't.
How Moamao implements this
Moamao packages this architecture as an API. One POST https://api.moamao.com/v1/runsfans your prompt out to a panel of up to four proposers across the claude, gpt, gemini, and llama families, runs a critique layer where each model reads the others' drafts, and has a judge model synthesize the final answer. Every draft, critique, and judge decision is recorded in a full trace you can replay. It's BYOK by default — model usage stays on your own provider keys, and Moamao charges a per-run orchestration fee — and smart routing sends prompts that don't need a panel to a single model, billed as a quarter run.