A curated list of tools and infrastructure for AI agent-to-agent communication, coordination, and collaboration.
Unlike general "awesome agents" lists that focus on what agents can do, this list focuses on how agents work together — the communication protocols, messaging systems, orchestration frameworks, shared memory layers, identity/trust mechanisms, and observability tools that make multi-agent systems possible.
2026 context: MCP has 97M+ downloads, A2A is adopted by 150+ organizations, and multi-agent systems are entering production at scale. This list tracks the infrastructure layer that makes it all work.
- Communication Protocols
- Messaging & Notification
- Multi-Agent Orchestration
- State & Memory
- Identity & Trust
- Observability & Debugging
- Developer Tools & SDKs
- Examples & Demos
- Learning Resources
Standards and protocols that define how agents discover, communicate, and delegate tasks to each other.
- MCP (Model Context Protocol) - Open standard by Anthropic for connecting AI models to tools and data sources. Governed by Linux Foundation. 97M+ downloads. The de facto standard for agent-to-tool communication.
- A2A (Agent-to-Agent Protocol) - Open protocol by Google (donated to Linux Foundation, June 2025) for peer-to-peer agent coordination. 150+ supporting organizations. Peer-to-peer agent communication and task delegation via Agent Cards.
- ACP (Agent Communication Protocol) - RESTful agent interoperability protocol originated by IBM/BeeAI, donated to Linux Foundation. Enables agents built with different frameworks (LangChain, CrewAI, AutoGen, etc.) to communicate via a standardized API.
- ANP (Agent Network Protocol) - Decentralized protocol for open agent networks, built on W3C DID standards. Enables agents to discover and interact without centralized coordination.
- UTCP (Universal Tool Calling Protocol) - Exposes tools to models using the tool's native endpoint rather than requiring an MCP wrapper. Contributed to Linux Foundation alongside ACP.
- OpenAI Agents SDK — Handoffs - First-class agent-to-agent handoff primitives in OpenAI's Agents SDK. Allows agents to delegate tasks with full context transfer.
Real-time messaging, pub/sub, event streaming, and notification infrastructure designed for or applicable to agent-to-agent communication.
- IM for Agents ⭐ - Agent-to-agent messaging without MCP. Three HTTP calls and your agents are talking — no SDK, no protocol implementation, no infrastructure setup. Cross-framework: Claude, GPT, Gemini, and local LLMs can all join the same room via REST API. Free tier available.
- im-agents-mcp - MCP server for IM for Agents. Allows any MCP-compatible agent to send and receive messages from other agents via IM for Agents.
- Agent Notify Action - GitHub Action for sending notifications from automated agents to humans or other systems. Supports Telegram, Slack, and webhooks.
- NATS - High-performance, cloud-native messaging system. Subjects and consumer groups make it excellent for agent event streaming and task routing. Supports request-reply patterns common in agent architectures.
- Redis Pub/Sub + Streams - Battle-tested message patterns. Redis Streams provide persistent, consumer-group-based messaging ideal for agent pipelines where replay and backpressure matter.
- RabbitMQ - Mature message broker with flexible routing (topics, fanout, headers). Useful for agent task queues with acknowledgment and retry semantics.
- Apache Kafka - Distributed event streaming platform. Best for high-throughput, durable agent event logs and audit trails in large-scale multi-agent deployments.
- XMTP - Decentralized, end-to-end encrypted messaging protocol. Used for agent-to-agent messaging with cryptographic identity, especially in Web3 agent contexts.
Frameworks for defining agent roles, task delegation, coordination logic, and workflow execution across multiple agents.
- CrewAI - Role-based multi-agent framework. Agents have roles, goals, and backstories. Simple API, production-ready, 30K+ GitHub stars.
- LangGraph - Graph-based agent orchestration by LangChain. Fine-grained control over agent state machines. Best for complex branching and human-in-the-loop flows.
- AutoGen - Microsoft's framework for conversational multi-agent workflows. Strong support for human-in-the-loop, code generation, and group chat patterns.
- Google ADK (Agent Development Kit) - Google's official SDK for building agents that work with Gemini and A2A protocol. Includes built-in support for multi-agent delegation.
- Anthropic Agent SDK - Official Anthropic toolkit for building Claude-powered agents with tool use, multi-turn conversations, and agent handoffs.
- OpenAI Agents SDK - OpenAI's production agent framework (evolved from Swarm). Built-in support for tool use, handoffs, guardrails, and tracing.
- Semantic Kernel - Microsoft's open-source SDK for building AI agents and copilots. Supports agent-to-agent patterns via the Process Framework.
- Swarms - Production-grade swarm orchestration. Supports hierarchical, sequential, concurrent, and mixture-of-agents patterns. Battle-tested for enterprise deployments.
- AgentScope - Alibaba's multi-agent platform. Provides MsgHub and pipeline abstractions for efficient message routing between agents.
- Mastra - TypeScript-first agent framework with memory, tool-calling, workflows, and RAG. Designed for building production AI applications with multi-agent capabilities.
- smolagents - Hugging Face's minimal, code-centric agent framework. Agents write and execute Python code to achieve goals. Simple integration into multi-agent pipelines.
- open-multi-agent - Production-grade multi-agent orchestration. Model-agnostic, supports team collaboration, task scheduling, and inter-agent communication with shared memory.
- PydanticAI - Production-grade agent framework from the Pydantic team. Type-safe tool calling, dependency injection, and structured output. Supports multi-agent orchestration via agent handoff primitives.
- LlamaIndex Workflows - Event-driven, async multi-agent orchestration built into LlamaIndex. Define agents as workflow steps with typed events — clean model for complex agent DAGs.
- Dify - Open-source LLM app platform with visual agent builder. Supports multi-agent workflows, tool integration, and knowledge retrieval. 60K+ GitHub stars.
Tools for persisting agent context, shared knowledge, cross-session memory, and distributed state management.
- Mem0 - Adaptive memory layer for AI agents. Intelligently compresses history into optimized memory representations. Claims up to 80% prompt token reduction. Lowest integration friction for standalone memory.
- Zep - Long-term memory store with temporal knowledge graph. Tracks how facts change over time. Integrates structured business data with conversational history. Best for temporal-aware production pipelines.
- Letta (formerly MemGPT) - Agents with OS-style virtual memory management. Self-editing memory architecture with in-context vs archival storage. Includes REST API and development environment.
- Cognee - Knowledge-graph-first memory for agents. Builds structured knowledge graphs from unstructured data for rich semantic retrieval.
- LangMem - LangChain's memory abstraction layer. Multiple memory types (episodic, semantic, procedural) with a unified API.
- Supermemory - Universal memory API for AI agents. Import from anywhere, query with semantic search. Designed as a "memory backbone" across multiple agents.
- Redis (State Management) - Fast in-memory store for short-term agent state, session data, and distributed locks. Essential for stateful multi-agent systems requiring sub-millisecond access.
- Chroma - Open-source embedding database. The most popular vector store for agent memory. Simple API, runs locally or in the cloud. pip install chromadb.
- Qdrant - High-performance vector database written in Rust. Excellent for production agent long-term memory with filtering, payload indexing, and sparse vector support.
- Weaviate - Open-source vector database with built-in vectorization. Modules for text, images, and multi-modal memory. Native GraphQL API.
- Pinecone - Managed vector database. Serverless option allows agents to query semantic memory with zero infrastructure. Widely used in production agent deployments.
- Haystack DocumentStore - Pluggable document store interface for Haystack agents. Supports Elasticsearch, OpenSearch, Weaviate, Pinecone, and more. Uniform API across backends.
Specifications and tools for establishing agent identity, authentication, and trust relationships between agents. The defining security challenge of 2026.
- A2A Agent Cards - A2A protocol standard for agent self-description. Agents publish capabilities, authentication requirements, and endpoints in a standardized JSON format. Enables dynamic agent discovery.
- SPIFFE / SPIRE - Production-grade workload identity framework. Each agent gets a cryptographically verifiable SVID (SPIFFE Verifiable Identity Document) — no shared secrets. SPIRE handles issuance, rotation, and revocation. The current best practice for secure agent-to-agent mTLS.
- WIMSE (Workload Identity and Minimal Secrets) - IETF working group standard for workload/agent identity in cloud environments. Combines SPIFFE, OAuth 2.0, and token exchange for agent delegation. First draft published 2025, actively developed 2026.
- OAuth 2.0 for AI Agents - Using OAuth 2.0 / DPoP for agent-to-agent authorization. Best practice for agents calling APIs or other agents on behalf of users. IETF draft (draft-klrc-aiagent-auth-00) published March 2026.
- DID (Decentralized Identifiers) - W3C standard for cryptographic agent identity without central authority. Used as the foundation for ANP and decentralized agent networks.
- WorkOS - Enterprise identity and auth platform increasingly used for AI agent authorization. Fine-grained permissions, audit logs, and organization-level access control — applies cleanly to multi-agent systems where agents act on behalf of users.
- Permit.io - Policy-based authorization for apps and agents. Define what agents are allowed to do via ReBAC/ABAC policies. Supports runtime policy updates without redeploy.
Tools for tracing, monitoring, evaluating, and debugging multi-agent systems in production.
- LangSmith - End-to-end observability platform for LLM/agent applications by LangChain. Tracing, evaluation, dataset management, and prompt experimentation.
- Langfuse - Open-source LLM observability (MIT license). Acquired by ClickHouse. 19 Fortune 50 clients. Full-featured: tracing, evaluations, prompt management, datasets.
- Arize Phoenix - Open-source LLM observability. Trace agent runs, visualize spans, and evaluate outputs locally or in the cloud. Built on OpenTelemetry.
- AgentOps - Agent-specific observability and testing. Session replays, cost tracking, failure analysis, and compliance monitoring. Native integrations with CrewAI, AutoGen, LangChain.
- Maxim AI - Agent quality and observability platform. Evaluation pipelines, red-teaming, and production monitoring for agentic workflows.
- OpenTelemetry - Vendor-neutral observability framework. Use with OpenLLMetry for LLM-specific spans and agent traces.
- Weights & Biases Weave - W&B's LLM observability platform. Trace, evaluate, and monitor agent runs with rich visualization.
- Traceloop / OpenLLMetry - OpenTelemetry-based LLM observability. Automatic instrumentation for LangChain, CrewAI, AutoGen, and 20+ frameworks. Sends traces to any OTel-compatible backend (Datadog, Grafana, etc.).
- Helicone - Open-source LLM observability and caching proxy. Drop-in for OpenAI/Anthropic/Gemini. Per-request latency, cost, and error tracking with user-level attribution.
- Honeyhive - Agent evaluation and monitoring platform. Dataset management, A/B testing for prompts, and production quality monitoring. Focus on multi-step agent workflows.
SDKs, CLIs, and utilities that accelerate building multi-agent systems.
- im-for-agents-python - Python SDK for IM for Agents. Send and receive messages between agents with minimal boilerplate.
- multi-agent-starter - Starter template for building multi-agent applications with MCP, A2A, and IM for Agents integration out of the box.
- Dapr - Distributed Application Runtime. Service invocation, pub/sub, state management, and actor pattern — all as portable building blocks for agent microservices.
- BeeAI Framework - IBM's open-source agent framework powering the ACP protocol. Production-ready with tool use, memory, and structured output.
- FastMCP - The fastest way to build MCP servers in Python. Decorator-based API similar to FastAPI. Used by tens of thousands of MCP server authors. Now part of the official MCP Python SDK.
- Composio - Managed tool integration platform for AI agents. 250+ pre-built integrations (GitHub, Slack, Notion, etc.) with auth management. Works with LangChain, CrewAI, AutoGen, and any agent framework.
- Portkey AI Gateway - Open-source AI gateway with fallbacks, retries, load balancing, and request routing across 200+ LLMs. Essential for multi-agent deployments that use multiple model providers.
- LiteLLM - Unified API for 100+ LLMs. Use the same code to call OpenAI, Anthropic, Gemini, Groq, Ollama. Simplifies multi-agent systems where different agents use different models.
- instructor - Structured output extraction from LLMs using Pydantic. Critical for agent-to-agent communication with typed message schemas — ensures agents receive parseable, validated data.
- Arcade AI - Managed tool execution platform for AI agents. Secure, sandboxed execution with auth flows for enterprise tools. Reduces the surface area for agent security incidents.
Reference implementations and worked examples of multi-agent collaboration patterns.
- a2a-examples - Practical examples of A2A protocol agent communication. Covers agent discovery, task delegation, and streaming responses.
- multi-agent-demo - End-to-end demo of multiple agents collaborating via IM for Agents and MCP. Shows orchestrator + specialist agent pattern.
- multi-agent-patterns - Common patterns for building multi-agent systems with code examples. Covers pipeline, fan-out, consensus, and critic patterns.
- Google A2A Samples - Official Google samples for A2A protocol. Best starting point for understanding Agent Cards and task lifecycle.
- AutoGen Examples - Microsoft's reference examples for AutoGen agent patterns including group chat and code execution.
- LangGraph Multi-Agent Examples - Official LangGraph examples covering supervisor, hierarchical, and collaborative multi-agent patterns with state management.
- CrewAI Examples - Official CrewAI examples including trip planner, stock analysis, and game builder crews. Best starting point for role-based multi-agent systems.
- IM for Agents — 3 curl demo - Minimal demo of agent-to-agent messaging with 3 HTTP calls — no SDK, no framework. Shows the baseline for any cross-framework agent coordination.
Articles, specs, and guides for understanding the multi-agent collaboration landscape.
- A2A Protocol Specification - Full A2A specification, concepts, and implementation guide.
- MCP Documentation - Official MCP docs with quickstarts, tutorials, and server implementation guides.
- MCP vs A2A: Complete Guide (2026) - Deep comparison of when to use MCP vs A2A vs ACP.
- AI Agent Protocol Ecosystem Map 2026 - Visual overview of the complete agent protocol landscape.
- Survey of Agent Interoperability Protocols - Academic survey covering MCP, ACP, A2A, and ANP.
- 120+ Agentic AI Tools Mapped (2026) - Comprehensive landscape map across 11 agent tool categories.
- 4 Patterns AI Agents Use to Coordinate - Practical guide: shared files, MCP, Kafka, and REST messaging — when each pattern breaks in production.
- The Agent Identity Playbook (2026) - How to think about non-human identity (NHI) for AI agents in enterprise environments.
- Securing AI Agents: The Defining Cybersecurity Challenge of 2026 - Bessemer VP analysis of the agent security landscape — permissions, secrets, and trust boundaries.
Submissions welcome! Please read the contribution guidelines first.
Only add tools that are:
- Actively maintained (last commit within 12 months, or commercially supported)
- Genuinely useful for agent-to-agent collaboration (not just "agent-adjacent")
- Open source, or have a meaningful free tier
Maintained by @masstensor · IM for Agents — messaging infrastructure for AI agents