Premium
Services
Premium

Compare 15 Agent Swarm Frameworks: OpenAI, Google, Microsoft

Hazal Şimşek
Hazal Şimşek
updated on Nov 5, 2026

Anthropic split its research assistant into a lead agent and a team of parallel subagents, reporting 90.2% better results on its internal evaluations. Yet, the system used about 15 times as many tokens.1 How much of that overhead one pays depends partly on the framework.

In our orchestration benchmark, we ran the same five-agent workflow with the same model and tools on four frameworks, and LangGraph finished 2.2 times faster than CrewAI because the frameworks handle tool calls and context differently.

Compare eleven frameworks based on their features and prices, and learn how the market has changed in a year:

Product
Maintainer
License
Languages
Pricing
Agency Swarm
VRSEN
MIT
Python 3.12+
Free; no hosted offering
Agent Squad
2FastLabs (originally AWS Labs)
Apache 2.0
Python, TypeScript, Swift
Free; Bedrock usage billed by AWS
AgentScope
Alibaba
Apache 2.0
Python, Java
Free; optional deployment on Alibaba Cloud billed by usage
CAMEL
CAMEL-AI.org
Apache 2.0
Python
Free; no hosted offering
CrewAI
CrewAI Inc.
MIT
Python
Free; AMP: free tier (50 runs/mo), Professional: $25/mo
Google ADK
Google
Apache 2.0
Python, TypeScript, Go, Java, Kotlin
Free; Agent Engine hosting usage-based
LangGraph
LangChain Inc.
MIT
Python, JS/TS
Free; hosting via LangSmith Plus: $39/seat/mo
Mastra
Mastra (Kepler Software)
Apache 2.0 (except ee/ code)
TypeScript
Free; platform: free tier, Teams: $250/mo
MetaGPT
FoundationAgents
MIT
Python
Free; hosted product Atoms (formerly MGX) priced separately
Microsoft Agent Framework
Microsoft
MIT
Python, .NET (Go in preview)
Free; Foundry hosting usage-based

What is an agent swarm?

An agent swarm is a setup where several AI agents work on one shared task, instead of a single agent doing everything (a “monolithic” design). The agents coordinate in one of three ways:

  • Handoffs: one agent passes control of the conversation to another, more specialized agent.
  • Message passing: agents send each other messages or results.
  • Shared state graph: agents are nodes in a graph that read and write a common state.

Agent swarm features

Agent swarm frameworks share four core features:

  • Tool and function calling: agents can call APIs and other functions, and the framework generates the schemas for those calls.
  • Multi-agent orchestration: done through explicit graphs, handoffs, hierarchical crews, or a central router that sends each task to the best-suited agent.
  • Memory and state: conversation history or task state carries over between turns, kept in memory or in a pluggable store such as SQLite, Redis, or a vector database.
  • Open source: every framework’s code is on GitHub under a permissive license, except Mastra.

Agent swarm frameworks and platforms

Agency Swarm

Built on top of the OpenAI Agents SDK. You define agents by role (for example CEO, developer or assistant) and declare which agents may contact which; agents then talk through a send_message tool. Tools are typed with Pydantic, and you save conversation history yourself through callback functions.

Pros:

  • Converts OpenAPI schemas directly into agent tools.
  • Runs on macOS, Linux and Windows.

Cons:

  • Models other than OpenAI’s (Anthropic, Gemini, Grok, Azure, OpenRouter) go through LiteLLM rather than a native integration.
  • Version 1.x is built on the OpenAI Agents SDK and Responses API, so changes there affect it.2

AgentScope

Alibaba’s Python framework, now at version 2.0, built from building blocks such as a ReAct loop, toolkit, memory and middleware. It requires Python 3.11 or higher and includes Agent Team, where a leader agent coordinates worker agents. Its permission system gives fine-grained control over tools and resources, including confirmation steps.

Pros:

  • Tools and code run in isolated workspaces on backends ranging from local and Docker to E2B, Daytona and Kubernetes.
  • Its agent service ships SQL and NoSQL persistence for agent state and sessions, plus scheduled tasks.3

Cons:

  • The quickstart example uses Alibaba’s DashScope model.3

Agent Squad

A classifier sends each request to the most suitable agent, using Amazon Bedrock by default.4 SupervisorAgent coordinates sub-agents that share one conversation context. GroundedAgent keeps data gathering separate from writing the answer, to reduce made-up content.

Pros:

  • The Swift version runs entirely on the device (iOS 16+, macOS 14+), including MCP tools, voice, and locally stored chat history.
  • Python and TypeScript have feature parity.
  • Ready-made agents cover Bedrock, Anthropic, OpenAI, Lex and Lambda.

Cons:

  • Routing adds a classifier call to each request; the faster JevClassifier option relies on a paid third-party service, TypeSafe AI.
  • GroundedAgent uses two LLM calls on turns that use tools.5

CAMEL

A Python framework from CAMEL-AI.org focused on finding the scaling laws of agents.Its Workforce module assembles, manages and scales teams of agents for tasks beyond a single agent, and it starts recovery steps when a task fails.

Pros:

  • Its design covers data generation, task automation and world simulation.
  • Agents can be given a HumanToolkit to ask a person for clarification or approval mid-task.

Cons:

  • It has no paid tier and no hosted service from CAMEL-AI.6
  • Tool sets such as web search are installed separately as extras (camel-ai[web_tools]).7

CrewAI

Agents are organized into “crews,” where each agent has a role, a goal, and tools. A second layer called Flows handles step-by-step, event-driven execution with conditional branches.

Memory is stored in a hierarchy, searchable with fast vector search or deeper LLM-assisted recall, and saves happen in the background. The hosted platform, AMP, adds a visual editor; AMP Factory runs the same platform on your own servers or private cloud.

Pros:

  • Standalone, with no LangChain dependency.
  • Supports MCP and A2A.
  • Memory and models can run locally through Ollama, so content doesn’t go to an outside API.8
  • Passes every earlier agent’s full output to the next one; in our orchestration benchmark, its final agent produced the most detailed result (5,339 tokens).

Cons:

  • Agents call the OpenAI API by default unless you configure another provider.
  • Anonymous usage telemetry is on by default (disable it with the OTEL_SDK_DISABLED environment variable).
  • Agents deliberate before calling a tool, which can’t be switched off; in our benchmark this took about 5 seconds, most of one agent’s 9-second run.

Google Agent Development Kit (ADK)

ADK 2.0 added a graph-based workflow engine that mixes fixed code steps with agent decisions, with support for routing, fan-out and fan-in, loops, retries and human-in-the-loop. The adk web interface includes a drag-and-drop Visual Builder, which is experimental and Python-only. Agents deploy to Vertex AI Agent Engine, Cloud Run or GKE.9

Pros:

  • Evaluation is built in through adk eval, rather than requiring a separate tool.
  • The Task API handles structured delegation between agents.

Cons:

  • The language versions don’t move together: 2.0 is released for Python, Go and TypeScript, but not yet for Java or Kotlin.

LangGraph

Workflows are defined as directed graphs that can loop and branch on conditions. An interrupt API lets a human inspect or change a run’s state at any point. LangSmith Studio is a visual debugger for graphs written in code. LangSmith Agent Builder configures agents through a form rather than a drag-and-drop canvas.10 LangSmith Deployment is LangChain’s optional managed hosting.

Pros:

  • Can be used without LangChain or any paid service.
  • Durable execution saves checkpoints, so a failed run resumes exactly where it stopped.
  • Passes only state changes between agents, not full outputs; in our orchestration benchmark, its final agent produced 2,589 tokens, the fewest of the frameworks tested.

Cons:

  • It describes itself as a low-level framework, and for building agents quickly its own README points users to Deep Agents, a higher-level package built on top of it.11

Mastra

Combines agents, graph-based workflows, memory, evaluations and observability in one framework. Workflows can pause indefinitely for human approval and resume from saved state. Studio is an interface for testing agents written in code, and it can be shared with a team through Mastra Cloud.12

Pros:

  • Connects to 40+ model providers through one interface.
  • Can publish its agents and tools as MCP servers.
  • Embeds in existing React, Next.js or Node.js apps.13

Cons:

  • Features in ee/ directories require a commercial license for production use.13

MetaGPT

A Python framework built on the idea that code is the result of applying standard operating procedures to a team of LLM agents. It assigns roles such as product manager, architect, project manager and engineer to separate agents.

Pros:

  • It runs as a CLI or a Python library, and the config accepts providers such as OpenAI, Azure, Ollama and Groq.
  • It includes a Data Interpreter role for writing code for data tasks.14

Cons:

  • The latest README news is from March 2025 and promotes mgx.dev, the team’s commercial product.14

Microsoft Agent Framework

An open, multi-language framework for building production-grade AI agents and multi-agent workflows. It is fully supported in both Python and C#/.NET with consistent APIs, and the README links migration guides from Semantic Kernel and from AutoGen.15

Pros:

  • Every orchestration supports human-in-the-loop, so a workflow can pause and wait for a person at set points.16
  • Agents can be deployed to Foundry-hosted infrastructure with two extra lines of code, and OpenTelemetry tracing is built in.15

Cons:

  • The Go version is a public preview that does not yet have declarative agents, RAG, CodeAct or functional workflows.
  • Some provider packages are still early. The Gemini integration is marked as Alpha.17

Microsoft AutoGen

Agents coordinate through a group chat rather than a defined execution graph. AutoGen Studio is a no-code visual builder, and UserProxyAgent provides human-in-the-loop input.18 Microsoft points new projects to Microsoft Agent Framework, and a community fork, AG2, continues separately.19

Pros:

  • A layered design offers both a high-level API for common patterns and a low-level core for custom ones, with Python and .NET interoperability.
  • Group chat suits multi-turn tasks such as brainstorming and iterative review.20

Cons:

  • The project is now community-managed, so support responses may be slow.
  • AutoGen Studio is meant for prototyping and, by its own README, is not a production-ready app.21

Microsoft Semantic Kernel

An SDK built around plugins, with connectors to existing enterprise systems and a Process Framework for structured business workflows. Microsoft provides migration guides that carry existing plugins over to Microsoft Agent Framework.22

Pros:

  • Model-agnostic, with MCP support for plugins.
  • Runs local models through Ollama, LM Studio or ONNX.

Cons:

  • Its README now states that Semantic Kernel “is now Microsoft Agent Framework,” and new features go there.

OpenAI Agents SDK

Built around agents, handoffs, guardrails, sessions and tracing. Sandbox agents work inside a container on long tasks and keep workspace state between steps, and realtime agents handle voice. It replaces OpenAI’s retired Swarm library.

Pros:

  • Human-in-the-loop is built in, and input and output guardrails can run in parallel and stop a run early.23
  • Works with over 100 models beyond OpenAI’s own.24

Cons:

  • The default local sandbox runs only on macOS and Linux; Windows needs Docker or a hosted sandbox.
  • Handoffs are the only coordination primitive, so a rule like “two agents must both finish before a third starts” has to be built outside the SDK.20

PraisonAI

Agents and multi-agent setups can be defined in YAML without code, built in the Flow drag-and-drop editor (based on Langflow), or written in Python or JavaScript. It ships with over 100 built-in tools and supports over 100 models.

Pros:

  • The Claw dashboard connects agents to Telegram, Discord, Slack and WhatsApp.
  • Doom-loop detection, human approval and guardrails help prevent runaway agents.

Cons:

  • Claw requires both OpenAI and Tavily API keys.
  • The project lives on its creator’s personal GitHub account.25

Spring AI Alibaba

The agent framework is a Java framework built on Spring AI with two layers. It includes built-in context engineering and human-in-the-loop support, along with workflow agents such as SequentialAgent, ParallelAgent, RoutingAgent and LoopAgent. The Graph runtime underneath provides persistence, orchestration and streaming for long-running stateful agents.26

Pros:

  • Agent-to-agent communication works through Nacos, so agents can coordinate across services.26
  • The Admin platform lets you build agents visually, deploy them without code or export them as a Java project, and it can migrate Dify DSL.26
  • It exclusively uses Java and is tightly integrated with Alibaba Cloud services such as Bailian, ARMS and Nacos.

Swarms

Through a SwarmRouter interface, you choose from ready-made coordination patterns such as sequential, concurrent, hierarchical, graph and mixture-of-agents. AutoSwarmBuilder generates an agent team from a task description, and consensus and debate patterns let agents vote. Swarms Cloud is a hosted API billed by usage.

Pros:

  • Documents ten architectures in its README and links to 60+ more.
  • GraphWorkflow runs independent branches in parallel automatically.
  • Claims compatibility with agents built in LangChain, AutoGen and CrewAI.

Cons:

  • The recommended “auto” mode has no fixed iteration cap, and the README itself advises a fixed cap for cost-sensitive work.27
Get our team to automate one of your business processes with AI agents, free of charge.
Automate a process

1. Major vendors replaced their first agent frameworks

Between March 2025 and April 2026, OpenAI, Microsoft and Google each replaced or rebuilt their agent frameworks.

  • OpenAI: Swarm was replaced by the Agents SDK in March 2025. The SDK keeps Swarm’s handoff model and adds AI guardrails, tracing and sessions.28
  • Microsoft: AutoGen and Semantic Kernel merged into Microsoft Agent Framework, which reached version 1.0 in April 2026. Both older frameworks are now in maintenance mode.29
  • Google: ADK 2.0 replaced its hierarchical agent executor with a graph-based workflow engine.30
  • AWS: Agent Squad left AWS Labs when its original authors took it to their own company, 2FastLabs.31

2. Graph-based workflow engines became standard

In 2026, ADK, Microsoft Agent Framework, LangGraph and Mastra all let developers build multi-agent workflows as graphs. Agents and functions are the nodes, and the edges set the order, branches and loops. 20 A graph can express any of the three common coordination patterns and most production systems combine more than one.

Centralized (orchestrator-worker)

A supervisor agent splits the task, sends subtasks to worker agents, and combines their results. Some examples are:

  • Amazon Bedrock: Multi-agent collaboration uses this model and has been generally available since March 10, 2025.32
  • Anthropic: Its research system uses a lead agent with parallel subagents. 1

Hierarchical

Supervisors delegate to lower-level supervisors, which helps with tasks too large for one agent’s context window.

  • ADK 2.0: The Workflow Runtime supports nested workflows, fan-out and fan-in, loops and retries.33
  • ADK templates: SequentialAgent, ParallelAgent and LoopAgent cover common pipelines without building a graph by hand.34

Decentralized (peer-to-peer handoffs)

Agents pass control directly to one another while the task runs, with no supervisor.

For example, LangGraph’s swarm library implements this and remembers the last active agent, so a follow-up message returns to it.35

Whichever pattern a framework uses, the handoff itself is rarely the bottleneck. In our orchestration benchmark, passing work from one agent to the next took milliseconds on every framework; the differences came from how each framework handles tool calls and context.

3. The two main agent protocols moved to the Linux Foundation in 2025

An agent can use both:

  • Model Context Protocol (MCP) connects an agent to its tools and data. Anthropic created it in 2024 and donated it to the Linux Foundation’s Agentic AI Foundation in December 2025.36
  • Agent2Agent (A2A) connects separate agents to each other, even across frameworks and vendors. Google created it in April 2025 and moved it to the Linux Foundation that June.37

Further reading

Don’t miss our benchmarks and data-driven insights. The button opens Google; selecting AIMultiple confirms that you wish to see AIMultiple more often in Google search results.
GoogleAdd as preferred source

Cite this research

Pick the format that matches where you're publishing. Pasting the link version into your CMS preserves the backlink.

Hazal Şimşek (2026) - "Compare 15 Agent Swarm Frameworks: OpenAI, Google, Microsoft". Published online at AIMultiple.com. Retrieved November 5, 2026, from: https://aimultiple.com/agent-swarm [Online Resource]

Şimşek, H. (2026, November 5). Compare 15 Agent Swarm Frameworks: OpenAI, Google, Microsoft. AIMultiple. https://aimultiple.com/agent-swarm

@misc{simsek2026,
  author = {Şimşek, Hazal},
  title  = {{Compare 15 Agent Swarm Frameworks: OpenAI, Google, Microsoft}},
  year   = {2026},
  month  = nov,
  howpublished    = {\url{https://aimultiple.com/agent-swarm}},
  note   = {AIMultiple. Retrieved November 5, 2026}
}
Download all data

Results and timestamps of 475 data points. Download the summary data shown in this article's charts and tables as a ZIP file containing 3 CSV files.

Last updated: October 28, 2026
Download

Want the granular data behind it? Join Premium

Reference Links

1.
How we built our multi-agent research system \ Anthropic
Anthropic
2.
GitHub - VRSEN/agency-swarm: Reliable Multi-Agent Orchestration Framework · GitHub
3.
GitHub - agentscope-ai/agentscope: Build and run agents you can see, understand and trust. · GitHub
4.
Agent Squad Documentation · DocSearch MCP
5.
GitHub - 2FastLabs/agent-squad: Flexible and powerful framework for managing multiple AI agents and handling complex conversations · GitHub
6.
CAMEL Review: Multi-Agent-Framework & Alternativen
PromptQuorum
7.
GitHub - camel-ai/camel: CAMEL: The first and the best multi-agent framework. Finding the Scaling Law of Agents. https://www.camel-ai.org · GitHub
8.
Memory - CrewAI
CrewAI
9.
Visual Builder - Agent Development Kit (ADK)
10.
Introducing LangSmith’s No Code Agent Builder
LangChain Blog
11.
GitHub - langchain-ai/langgraph: Build resilient agents. · GitHub
12.
Introducing Mastra’s Agent Studio | Mastra Blog
Mastra
13.
GitHub - mastra-ai/mastra: Mastra is the modern TypeScript framework for AI-powered applications and agents. · GitHub
14.
GitHub - FoundationAgents/MetaGPT: The Multi-Agent Framework: First AI Software Company, Towards Natural Language Programming · GitHub
15.
GitHub - microsoft/agent-framework: A framework for building, orchestrating and deploying AI agents and multi-agent workflows with support for Python and .NET. · GitHub
16.
Microsoft Agent Framework Workflows - Human-in-the-loop (HITL) | Microsoft Learn
17.
https://raw.githubusercontent.com/microsoft/agent-framework/main/python/packages/gemini/pyproject.toml
18.
| AutoGen for .NET
19.
AutoGen Explained: Status, Architecture and Alternatives [2026]
Atlan
20.
LLM-Based Multi-Agent Orchestration: A Survey of Frameworks, Communication Protocols, and Emerging Patterns
MDPI
21.
GitHub - microsoft/autogen: A programming framework for agentic AI · GitHub
22.
Semantic Kernel to Microsoft Agent Framework Migration Guide | Microsoft Learn
23.
OpenAI Agents SDK
24.
GitHub - openai/openai-agents-python: A lightweight, powerful framework for multi-agent workflows · GitHub
25.
GitHub - MervinPraison/PraisonAI: PraisonAI — Hire a 24/7 AI Workforce. Stop writing boilerplate and start shipping autonomous self-improving agents that research, plan, code, and execute tasks. Deployed in 5 lines of code with built-in memory, RAG, and
26.
GitHub - alibaba/spring-ai-alibaba: Agentic AI Framework for Java Developers · GitHub
27.
GitHub - kyegomez/swarms: The Enterprise-Grade Multi-Agent Orchestration Framework. Website: https://swarms.ai · GitHub
28.
GitHub - openai/swarm: Educational framework exploring ergonomic, lightweight multi-agent orchestration. Managed by OpenAI Solution team. · GitHub
29.
Microsoft Agent Framework Version 1.0 | Microsoft Agent Framework
30.
Welcome to ADK 2.0 - Agent Development Kit (ADK)
31.
agent-squad/README.md at main · 2FastLabs/agent-squad · GitHub
32.
Amazon Bedrock announces general availability of multi-agent collaboration | Artificial Intelligence
33.
Graph-based agent workflows - Agent Development Kit (ADK)
34.
Parallel workflow - Agent Development Kit (ADK)
35.
GitHub - langchain-ai/langgraph-swarm-py: For your multi-agent needs · GitHub
36.
MCP joins the Agentic AI Foundation | Model Context Protocol Blog
37.
GitHub - a2aproject/A2A: Agent2Agent (A2A) is an open protocol enabling communication and interoperability between opaque agentic applications. · GitHub
Hazal Şimşek
Hazal Şimşek
Industry Analyst
Hazal is an industry analyst at AIMultiple, focusing on process intelligence (including process mining) and enterprise automation (including IT automation and low-code/no-code automation).
View Full Profile

Be the first to comment

Your email address will not be published. All fields are required. Comments are left in their original language.

0/450