Awesome Claude Multi-Agent

A curated list of frameworks, patterns, protocols, and case studies for multi-agent orchestration where Claude is a primary model.

Multi-agent here means coordinated systems of multiple agents across processes, not single-session Claude Code subagents.

Companion documents: glossary for terminology, pitfalls for failure modes, and security references.

Contents

• Foundational Reading
• Multi-Agent Patterns and Architectures
• Orchestration Frameworks
• Subagent Collections
• Starter Templates and Examples
• Skills, Hooks, and MCP for Coordination
• Memory, State, and Communication
• Standards and Protocols
• Evaluation and Benchmarks
• Datasets and Environments
• Observability and Debugging
• Security, Safety, and Trust
• Research Papers
• Case Studies and Production Postmortems
• Talks, Courses, and Books
• Community

Foundational Reading

Primary sources from Anthropic and seminal multi-agent LLM papers that establish the concepts behind this list. Listed chronologically.

• ReAct: Synergizing Reasoning and Acting in Language Models - Interleaves reasoning traces with tool actions; baseline for tool-using agent orchestration (ICLR 2023).
• Reflexion: Language Agents with Verbal Reinforcement Learning - Verbal self-reflection between attempts lets an actor agent learn from prior failures across episodes (NeurIPS 2023).
• CAMEL: Communicative Agents for Mind Exploration - Cooperative role-playing via inception prompting; first scalable two-agent coordination paper (NeurIPS 2023).
• HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face - LLM controller plans tasks then delegates to specialist Hugging Face models; canonical orchestrator-of-models pattern (NeurIPS 2023).
• Generative Agents: Interactive Simulacra of Human Behavior - Twenty-five LLM agents simulate daily life with memory, reflection, and planning (UIST 2023).
• Self-Refine: Iterative Refinement with Self-Feedback - Generator, critic, and refiner roles within one model; defines the reflection pattern (NeurIPS 2023).
• Improving Factuality and Reasoning via Multiagent Debate - Multiple LLM instances debate over rounds to improve factuality and reasoning (ICML 2024).
• MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework - SOP-encoded multi-agent software company with product manager, architect, and engineer roles (ICLR 2024).
• AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation - Framework for conversable multi-agent chat with customizable roles and tool access (ICLR 2024).
• Mixture-of-Agents Enhances Large Language Model Capabilities - Layered LLM ensemble where each layer's agents read the previous layer's outputs (2024).
• Building Effective Agents - Anthropic's reference patterns for agents: prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer (2024).
• - Anthropic's engineering account of running parallel Claude orchestrator-worker research agents (2025).

Multi-Agent Patterns and Architectures

Canonical coordination patterns with implementation references. See Foundational Reading above for the seminal papers.

• Blackboard - Shared workspace where agents with varied roles read and write to a common store until consensus is reached.
• Dynamic Agent Selection - Agents are chosen and wired into a communication structure on the fly per task; defined by the DyLAN network.
• Hierarchical Role-Based - Structured agent hierarchy with defined SOPs per role; MetaGPT is the canonical implementation.
• Market-Based Coordination - Agents act as market participants trading probabilistic beliefs to converge on shared, truthful outcomes.
• Mixture-of-Agents - Layered aggregation where each layer's agents improve on the previous layer's proposals.
• Multi-Agent Debate - Multiple instances independently reason then critique each other over iterative rounds.
• Orchestrator-Worker - Lead agent plans and spawns parallel subagents; defined in Anthropic's agent design patterns guide.
• Parallelization and Voting - Independent agents answer the same query and a majority vote selects the result; accuracy scales with agent count.
• Recursive Task Decomposition - Tasks are decomposed into subtasks only when an agent cannot execute them directly; defined by ADaPT.
• Reflection and Self-Critique - Actor produces output; critic proposes revisions; actor refines; loop until quality gate.
• Sequential Handoff - One agent transfers control and context to a more specialized agent mid-conversation.
• Supervisor with Tool Routing - Supervisor node routes to specialized agents via tool calls; reference implementation in LangGraph.

Orchestration Frameworks

Tools and libraries that coordinate multiple Claude instances or Claude-compatible agents across processes. Listed alphabetically.

• AG2 - Apache-licensed AutoGen community fork providing multi-agent AgentOS primitives and active development.
• AgentScope - Alibaba's distributed multi-agent platform with message-passing primitives, fault tolerance, and visual workflow studio.
• Agno - High-performance Python framework for multi-agent teams with memory and knowledge; model-agnostic interface including Claude.
• Atomic Agents - Schema-driven framework for building and chaining modular agents into multi-agent pipelines; Claude via Instructor.
• AutoGen - Microsoft's conversable multi-agent framework; Claude supported via a custom model client extension.
• BeeAI Framework - IBM-originated Python and TypeScript framework for agent workflows and multi-agent orchestration across model providers.
• CAMEL - Research framework for role-playing multi-agent systems with stateful memory and Claude support.
• ChatDev - Chat-chain platform where agent teams with distinct roles build software end-to-end.
• Claude Agent SDK - Anthropic's official Python SDK for building Claude agents with tools, hooks, and MCP servers.
• Claude Squad - Terminal manager for running multiple parallel Claude Code agents in isolated Git workspaces.
• Claude Swarm - Decomposes tasks across parallel Claude Code agents with conflict detection and a rich terminal coordination UI.
• Conductor - Microsoft CLI for defining and running multi-agent workflows on the Anthropic Agents SDK and GitHub Copilot SDK.
• CrewAI - Role-based multi-agent framework with first-class Claude support, crew primitives, and shared memory.
• - Framework for resilient, stateful multi-agent systems with workflow orchestration on the Dapr runtime.