Compare 12 Agent Swarm Frameworks: OpenAI, Google, Microsoft

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:

What is an agent swarm?

An agent swarm is a setup where several KI 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

Agent Squad

A classifier sends each request to the most suitable agent, using Amazon Bedrock by default.³ 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 KI.
• GroundedAgent uses two LLM calls on turns that use tools.⁴

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.⁵

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.⁶
• 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.⁷

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.⁸ LangSmith Agent Builder configures agents through a form rather than a drag-and-drop canvas.⁹ 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.¹⁰

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.¹¹

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.¹²

Cons:

• Features in ee/ directories require a commercial license for production use.¹²

Microsoft Agent Framework

An open, multi-language framework for building production-grade KI 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.¹³
Pros:

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

Cons:

• Some provider packages are still early ¹⁴

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.¹⁵ Microsoft points new projects to Microsoft Agent Framework, and a community fork, AG2, continues separately.¹⁶

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.¹⁷

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.¹⁸

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.¹⁹

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.²⁰
• Works with over 100 models beyond OpenAI’s own.²¹

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.¹⁷

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.²²

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 README makes enterprise claims that can’t be checked for a self-hosted library, such as a 99,9 %+ uptime guarantee.
• The recommended “automatisch” mode has no fixed iteration cap, and the README itself advises a fixed cap for cost-sensitive work.²³

Agent swarm orchestration patterns and trends

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 KI guardrails, tracing and sessions.²⁴
• 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.²⁵
• Google: ADK 2.0 replaced its hierarchical agent executor with a graph-based workflow engine.²⁶
• AWS: Agent Squad left AWS Labs when its original authors took it to their own company, 2FastLabs.²⁷

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. ¹⁷ 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, 2025.²⁸
• 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.²⁹
• ADK templates: SequentialAgent, ParallelAgent and LoopAgent cover common pipelines without building a graph by hand.³⁰

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.³¹

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 2025.³²
• 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.³³

Further reading

• Major Agentic AI Design Patterns & Real-World Examples
• LLM Orchestration: Top Frameworks and Gateways
• Top KI Agent Builders

Verpassen Sie nicht unsere Benchmarks und datengestützten Erkenntnisse. Die Schaltfläche öffnet Google; die Auswahl von AIMultiple bestätigt, dass Sie AIMultiple häufiger in den Google-Suchergebnissen sehen möchten.

Als bevorzugte Quelle hinzufügen

Diese Forschung zitieren

Wählen Sie das Format, das zu Ihrem Veröffentlichungsort passt. Wenn Sie die Link-Version in Ihr CMS einfügen, bleibt der Backlink erhalten.

@misc{simsek2026,
  author = {Şimşek, Hazal},
  title  = {{Compare 12 Agent Swarm Frameworks: OpenAI, Google, Microsoft}},
  year   = {2026},
  month  = oct,
  howpublished    = {\url{https://aimultiple.com/agent-swarm}},
  note   = {AIMultiple. Abgerufen am 24. Oktober 2026}
}

Referenzlinks

1.

How we built our multi-agent research system \ Anthropic

Anthropic

2.

GitHub - VRSEN/agency-swarm: Reliable Multi-Agent Orchestration Framework · GitHub

3.

https://docsearch.algolia.com/mcp/docs/repo/2fastlabs/agent-squad

4.

GitHub - 2FastLabs/agent-squad: Flexible and powerful framework for managing multiple AI agents and handling complex conversations · GitHub

5.

https://www.crewai.com/pricing

6.

Memory - CrewAI

Mintlify

7.

https://adk.dev/visual-builder/

8.

LangSmith Studio - Docs by LangChain

Docs by LangChain

9.

https://www.langchain.com/blog/langsmith-agent-builder

10.

GitHub - langchain-ai/langgraph: Build resilient agents. · GitHub

11.

https://mastra.ai/blog/agent-studio

12.

https://github.com/mastra-ai/mastra

13.

README

14.

githubusercontent

15.

| AutoGen for .NET

16.

https://atlan.com/know/ai-agent/what-is-autogen/

17.

LLM-Based Multi-Agent Orchestration: A Survey of Frameworks, Communication Protocols, and Emerging Patterns

MDPI

18.

https://github.com/microsoft/autogen

19.

Semantic Kernel to Microsoft Agent Framework Migration Guide | Microsoft Learn

20.

OpenAI Agents SDK

21.

GitHub - openai/openai-agents-python: A lightweight, powerful framework for multi-agent workflows · GitHub

22.

https://github.com/MervinPraison/PraisonAI

23.

GitHub - kyegomez/swarms: The Enterprise-Grade Multi-Agent Orchestration Framework. Website: https://swarms.ai · GitHub

24.

GitHub - openai/swarm: Educational framework exploring ergonomic, lightweight multi-agent orchestration. Managed by OpenAI Solution team. · GitHub

25.

https://devblogs.microsoft.com/agent-framework/microsoft-agent-framework-version-1-0/

26.

https://adk.dev/2.0/

27.

Rate limit · GitHub

28.

Amazon Bedrock announces general availability of multi-agent collaboration | Artificial Intelligence

29.

https://adk.dev/graphs/

30.

https://adk.dev/agents/workflow-agents/parallel-agents/

31.

GitHub - langchain-ai/langgraph-swarm-py: For your multi-agent needs · GitHub

32.

https://blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation/

33.

GitHub - a2aproject/A2A: Agent2Agent (A2A) is an open protocol enabling communication and interoperability between opaque agentic applications. · GitHub

Hazal Şimşek

Branchenanalystin

Folgen auf

Hazal ist Branchenanalystin bei AIMultiple und konzentriert sich auf Prozessintelligenz (einschließlich Process Mining) und Unternehmensautomatisierung (einschließlich IT-Automatisierung und Low-Code-/No-Code-Automatisierung).

Vollständiges Profil anzeigen

Seien Sie der Erste, der kommentiert

Ihre E-Mail-Adresse wird nicht veröffentlicht. Alle Felder sind erforderlich. Kommentare werden in ihrer Originalsprache belassen.

Name

E-Mail-Adresse

Kommentar

0/450

Kommentar posten