Mixture of Agents (MoA) routes queries to specialized agents based on task type, combining expert capabilities. Architecture: Router/gate model classifies query → selects appropriate specialist(s) → aggregates responses. Similarity to MoE: Like Mixture of Experts but at agent level rather than neural network layer. Routing strategies: Hard routing (one agent), soft routing (weighted combination), top-k (multiple specialists), learned routing function. Specialist types: Domain experts (coding, writing, analysis), task experts (search, calculation, planning), format experts (JSON, markdown, code). Router training: Classification on task types, learned from interaction data, or rule-based heuristics. Benefits: Specialized agents outperform generalists, efficient resource use, modular updates. Implementation: Query embedding → router model → agent selection → execution → response merging. Aggregation: Single response pass-through, synthesis across specialists, quality-based selection. Frameworks: LangChain routers, custom MoA implementations. Challenges: Routing accuracy, handling ambiguous queries, load balancing, maintaining consistency. Optimization: Cache routing decisions, batch similar queries, precompute agent capabilities.