In 2026 the multi-agent field has consolidated around a handful of production frameworks. By every adoption signal (package downloads, GitHub activity, enterprise case studies), the systems people actually use today are:
Two older names you may know have been retired: AutoGen (60.8k stars) is in maintenance mode, and OpenAI Swarm was an educational project replaced by the Agents SDK. Both point their users at their successors.
| Downloads/month | 62.6M | GitHub stars | ~134k (langchain monorepo) |
|---|
LangGraph is a low-level orchestration runtime that models multi-agent systems as stateful graphs with loops, persistence (checkpointing), and human-in-the-loop control. It hit a stable 1.x line and pairs with LangChain (1,000+ integrations), Deep Agents (a harness for long-running research/coding agents), and LangSmith for tracing and evaluation. It is the framework most production teams name first, and the one competitors benchmark against.
| Downloads/month | 36.6M | GitHub stars | 7.1k+ (young repo, split across SDKs) |
|---|
Open-source SDK "built from production systems inside Amazon", in Python and TypeScript. Model- and cloud-agnostic, with a model-driven design (the LLM plans; you supply tools, hooks, guardrails, and steering policies) and built-in multi-agent patterns such as Agent-as-Tool and Swarm. Deploys to Bedrock AgentCore, Lambda, Fargate, and EKS. Named production users include Smartsheet, Swisscom, Verisk, Eightcap, Zafran, Jit, and Tavily. Its download volume is amplified by AWS's own CI ecosystem, but the customer list is real.
| Downloads/month | 32.4M | GitHub stars | 29.2k (openai-agents-python) |
|---|
A deliberately small framework for multi-agent workflows: agents, tools, handoffs (delegation between agents), guardrails, and built-in tracing. It replaced the educational Swarm project (21.9k stars, now frozen). Works best on OpenAI models but supports others via LiteLLM. Teams typically add Temporal or DBOS for durable, long-running execution.
| Downloads/month | 31.3M | Languages | Python, TypeScript |
|---|
The agent loop that powers Claude Code, packaged as a library. Subagents are a first-class primitive ("spawn specialized agents for focused subtasks"), alongside hooks, MCP, permissions, sessions, skills, and plugins. This is arguably the most-used multi-agent system in day-to-day practice, because every Claude Code user who runs parallel subagents is running a multi-agent system, whether they call it that or not. Anthropic also offers hosted Managed Agents.
| Downloads/month | 25.8M (≈2.3M organic) | GitHub stars | ~49k |
|---|
The leading independent framework, built on a role-based mental model: each agent has a persona, goal, and tools, and a "crew" divides a task among them. Self-contained (no LangChain dependency), MIT-licensed, with MCP and A2A support. The company behind it now sells an enterprise "agent build & runtime" platform and claims usage by 65% of the Fortune 500, with case studies from Docusign, General Assembly, Gelato, Piracanjuba, and Konecta. Community-reported rough edges: non-OpenAI provider reliability and async execution.
| Downloads/month | 19.1M | GitHub stars | ~19k (adk-python) |
|---|
Google's opinionated, batteries-included framework, now shipping in Python, TypeScript, Go, Java, and Kotlin; ADK 2.0 added graph workflows. Comes with a CLI (adk web, adk run, adk api_server), a browser debugging UI, built-in sessions/memory, evals, and native MCP, A2A, and OpenAPI tool support. Strongest for teams deploying on Vertex AI Agent Engine, Cloud Run, or GKE.
| Downloads/month | 1.2M (agent-framework) | GitHub stars | 13.3k, very active |
|---|
The unified successor to AutoGen (60.8k stars, maintenance mode since its 1.0 GA) and Semantic Kernel (2.1M downloads/month, security-fixes-only). It merges AutoGen's conversational multi-agent patterns (sequential, concurrent, handoff, group chat, Magentic-One) with Semantic Kernel's enterprise features, ships Python and .NET runtimes simultaneously, and integrates Azure AI Foundry guardrails, OpenTelemetry, MCP, and A2A. Raw downloads are still small because it is young and most Microsoft-stack teams are mid-migration; the direction of travel is unambiguous, since Microsoft has pointed both predecessor communities at it.
| System | Signal | What it is |
|---|---|---|
| Pydantic AI | 10.7M dl/mo | Type-safe Python agent framework from the Pydantic team; growing fast as the "FastAPI of agents". |
| LlamaIndex Workflows | 6.3M dl/mo (llama-index) | Event-driven multi-agent orchestration for document-heavy pipelines. |
| Haystack (deepset) | 26.4k stars | Pipeline-based orchestration, strong in RAG-centric production systems. |
| Mastra | ~23k stars | TypeScript-first agent framework (from the Gatsby team): workflows, memory, Studio. |
| MetaGPT | 70.2k stars | "AI software company" simulation; hugely starred, research-flavored, lighter production use. |
| CAMEL | 17.7k stars | Academic multi-agent framework focused on agent-society scaling laws. |
| TradingAgents | 103k stars | Multi-agent LLM trading framework; the most-starred "multi-agent" repo on GitHub, popular with retail quants. |
| smolagents (Hugging Face) | 571k dl/mo | Minimal code-acting agents; popular for learning and lightweight tasks. |
| AG2 | 312k dl/mo | Community fork continuing the original AutoGen line. |
llama-agents is the maintained path| If you… | Use |
|---|---|
| Need precise control over stateful, cyclic multi-agent graphs with persistence and human-in-the-loop | LangGraph |
| Run on AWS / Bedrock and want enterprise guardrails with minimal code | Strands Agents |
| Want the simplest possible handoff-style delegation on OpenAI models | OpenAI Agents SDK |
| Are building coding/research agents, or want Claude Code's loop with subagents in your own app | Claude Agent SDK |
| Think in terms of role-based teams and want the fastest prototype (or a no-code enterprise platform) | CrewAI |
| Are GCP-native or need Go/Java/Kotlin SDKs | Google ADK |
| Are on Azure/.NET or migrating from AutoGen/Semantic Kernel | Microsoft Agent Framework |
| Want type-safe, Pythonic agents without heavy abstractions | Pydantic AI |
| Build TypeScript/Next.js products | Mastra (or Strands TS / ADK TS) |