moa-solvelisted

# MOA-Solve — Mixture-of-Agents for hard problems ## What this is A control plane for solving hard problems with multiple models, built on **Hermes' own MoA runtime**. Grounded in Mixture-of-Agents (Wang et al., Together AI, arXiv:2406.04692, ICLR 2025): a **reference layer** (several model families answer independently) plus an **aggregator** (one strong model writes a NEW synthesized answer). The paper's load-bearing result: the aggregator that _synthesizes_ beats an LLM-ranker that merely _picks the best proposal_. That is the empirical basis for "pull the best from EACH, don't pick the single best ONE." Important reframing (do not skip): "best from each" does NOT mean cut-and-paste splicing spans from references. Splicing produces Frankenstein artifacts where model A's architecture assumes state that model B's grafted completion never provides. It means: **synthesize a superior answer that is accountable to the strongest element found in each reference**, grafted through explicit interface checks. The component ledger (below) is how you do that without averaging to mush. **This skill does not implement fan-out.** Hermes already ships a real MoA runtime with parallel reference dispatch, per-slot provider/model routing, credential handling, prompt-cache decoration, usage/cost accounting, and trace persistence. This skill is the **method** — when to convene a panel, how to brief it, and how to synthesize the result. Do not hand-roll an HTTP fan-out script; see "Why there