- Home
- AI agents
- Platforms and frameworks
- Agent frameworks compared
Build · Platforms and frameworks
AI agent frameworks compared: which one should you build on?
An agent framework is a code library that runs the agent loop for you: it calls the model, executes tools, keeps state between steps and coordinates several agents. LangGraph, Microsoft Agent Framework, Strands Agents, Google ADK and the OpenAI and Claude agent SDKs are the main ones, and the cloud platforms now host them.
In this guide: 83 pages
Key takeaways
- Pick the framework for where it will run and who will maintain it. Foundry hosted agents and AgentCore Runtime both accept several frameworks, so the cloud does not force the choice.
- Churn is real. AutoGen is archived, Semantic Kernel is now Microsoft Agent Framework, and LangChain moved its prebuilt agent into langchain.agents, all within about a year.
- Keep agent logic in your own code and use MCP, A2A and OpenTelemetry at the edges. Those edges are what you can move when a framework or a platform changes.
- A managed harness is the alternative to a framework: you declare the model, tools and instructions and the vendor runs the loop, trading control for less code.
Which agent frameworks matter in 2026, and what changed?
This page compares eight widely used frameworks. Most reached a stable major version between mid-2025 and mid-2026, while the OpenAI Agents SDK is still on 0.x releases (Python 0.24.0 on 15 Oct 2026) and we did not verify CrewAI’s current version. The bigger change is consolidation: Microsoft folded its older frameworks into one, and each model vendor now pairs its SDK with a managed, hosted option.
Microsoft announced Agent Framework on 1 Oct 2025 and shipped 1.0 for .NET and Python on 3 Apr 2026; Go is in public preview. The AutoGen repository now reads “Project archived and no longer maintained”, and the Semantic Kernel README says Semantic Kernel is now Microsoft Agent Framework. LangChain and LangGraph reached 1.0 on 22 Oct 2025 with a promise of no breaking changes until 2.0; LangChain’s create_agent now runs on the LangGraph runtime, and langgraph.prebuilt is deprecated in favour of langchain.agents.
AWS released Strands Agents 1.0 on 15 Jul 2025 and uses it to power the AgentCore harness. Google open-sourced the Agent Development Kit on 9 Apr 2025 and made ADK Python 2.0, with a graph-based workflow runtime, generally available on 19 May 2026. OpenAI’s Agents SDK, the successor to its experimental Swarm, arrived on 11 Mar 2025, and Anthropic renamed the Claude Code SDK to the Claude Agent SDK on 29 Sep 2025. LlamaIndex Workflows reached 1.0 on 30 Jun 2025 as standalone packages.
How do the main agent frameworks compare?
They differ less in what they can do than in how they model orchestration, where state lives and which platforms document hosting them. The table compares them on those points as of 2 Nov 2026; the last column is our view of where each one fits.
Agent frameworks compared on orchestration, managed hosting and fit, as documented on 2 Nov 2026.
| Framework | Orchestration and state | Managed hosting documented | Fits when |
|---|---|---|---|
| LangGraph | Graph of nodes and edges; checkpointers keyed by thread_id; interrupts for approval | Foundry hosted agents; AgentCore Runtime | You want explicit control of every path in code you own |
| Microsoft Agent Framework | Graph workflows plus sequential, concurrent, handoff, group chat and Magentic patterns; checkpointing, pause and resume; Durable Extension (preview at 1.0) | Foundry hosted agents; Azure Functions through the Durable Extension | Your team works in .NET or Python on Azure and needs workflows with approvals |
| Strands Agents | Agents as tools, swarm, graph and workflow; sessions on filesystem or S3; A2A | AgentCore Runtime; the engine under the AgentCore harness | You build on AWS and may move from code to the managed harness later |
| Google ADK | Sequential, parallel and loop workflow agents; LLM transfer; 2.0 graph workflows with human input; A2A | Agent Runtime with Sessions and Memory Bank; AgentCore Runtime | You build on Google Cloud or want workflow agents in several languages |
| OpenAI Agents SDK | Agents, handoffs, agents as tools, guardrails; sessions; human in the loop; sandbox agents | Foundry hosted agents; AgentCore Runtime; Temporal for durability | You want a small API with handoffs, paired with a durable engine for long runs |
| Claude Agent SDK | The Claude Code agent loop with subagents, compaction and MCP | Foundry hosted agents; Claude Managed Agents as the hosted counterpart (beta) | Coding, research and file-heavy agents on Claude models |
| CrewAI | Role-based crews inside Flows with state, events and loops | AgentCore Runtime; CrewAI AMP for managed deployment and a no-code studio | Teams that think in roles and want a studio next to code |
| LlamaIndex Workflows | Event-driven steps; AgentWorkflow for several agents; a guide to durable workflows with checkpoints | AgentCore Runtime | Agents built around document parsing and retrieval |
How do you choose an agent framework?
Start from where the agent will run, who will maintain it and how long it must keep state. The frameworks overlap so much that those facts decide more than any feature list.
- Hosting first. Foundry hosted agents take Agent Framework, LangGraph, the OpenAI Agents SDK, the Anthropic agent SDK, the Copilot SDK or your own code as a container or .zip. AgentCore Runtime works with CrewAI, LangGraph, LlamaIndex, ADK, the OpenAI Agents SDK, Strands and any model. Check the target platform’s list before you write code.
- Language and team. Agent Framework ships for .NET and Python; Strands and the OpenAI Agents SDK have Python and TypeScript versions; ADK lists Python, TypeScript, Go, Java and Kotlin. Choose a language your platform team already supports in production.
- State and duration. If runs last minutes, any framework’s session store will do. If they last hours or wait for people, choose one with checkpoints and pair it with a durable engine: LangGraph checkpointers, the Agent Framework Durable Extension, or Temporal for the OpenAI Agents SDK.
- Telemetry. Agent Framework emits OpenTelemetry GenAI spans for agent, model and tool calls, with sensitive data off by default. Confirm that any framework you pick emits those conventions, which are still in Development status, so traces from every agent land in one view.
- Exit plan. Keep prompts, tool definitions and evaluation sets out of framework-specific formats, describe tools in MCP and use A2A between agents. A framework change then rewrites the loop and leaves the rest alone.
Do you need a framework, or will a managed harness do?
A managed harness is enough when the agent is one loop with tools and you accept the vendor’s defaults. A framework earns its place when you need custom orchestration, unusual state, or your own mix of models and clouds.
Every vendor now sells the loop as a service. The AgentCore harness, generally available since 17 Jun 2026, lets you declare the model, tools, skills and instructions while AgentCore handles the environment, compute, memory, identity, networking and observability, and AWS recommends it in place of Bedrock Agents Classic. Copilot Studio runs agents on three harnesses, Foundry offers prompt agents with an optional Microsoft Copilot harness in preview, Claude Managed Agents has been in beta since 8 Apr 2026, the OpenAI Agents API has been in public beta since 10 Sep 2026, and Google’s Managed Agents API entered preview on 19 May 2026.
The trade-off is control. AWS’s own comparison says routing-mode multi-agent “is not straightforward today” on the harness and that full multi-agent collaboration “requires custom framework code”. A harness also ties the agent loop to one vendor’s release schedule; OpenAI’s visual Agent Builder, launched in Oct 2025, shuts down on 30 Nov 2026. Start on a harness for a single-purpose agent, and move to a framework when you hit a limit you can name.
How do Microsoft, AWS and Google implement it?
Each cloud’s own framework, what it hosts and its managed harness, with statuses as the vendor pages showed on 2 Nov 2026.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Vendor framework | Microsoft Agent Framework 1.0 (GA 3 Apr 2026), which replaces Semantic Kernel and AutoGen | Strands Agents 1.0 (15 Jul 2025) | Agent Development Kit, Python 2.0 GA 19 May 2026 |
| Hosts other frameworks | Foundry hosted agents: Agent Framework, LangGraph, OpenAI Agents SDK, Anthropic agent SDK, Copilot SDK or custom code | AgentCore Runtime: CrewAI, LangGraph, LlamaIndex, ADK, OpenAI Agents SDK, Strands and any model | Agent Runtime, with ADK as the native framework |
| Managed harness | Copilot Studio harnesses; Foundry prompt agents with an optional Copilot harness (preview) | AgentCore harness, GA 17 Jun 2026 | Managed Agents API, preview since 19 May 2026 |
| Durable state | Agent Framework checkpoints; Durable Extension (preview at 1.0) | Strands session managers; Lambda durable functions; Step Functions | ADK sessions and Memory Bank on Agent Runtime |
LangGraph, CrewAI and LlamaIndex are independent of the clouds. OpenAI and Anthropic pair their SDKs with hosted options, the Agents API and Claude Managed Agents, that were still in beta on 2 Nov 2026.
Frequently asked questions
Which AI agent framework should we choose?
It depends on where the agent runs. On Azure with .NET or Python teams, Microsoft Agent Framework is the natural default; on AWS, Strands leads into the AgentCore harness; on Google Cloud, ADK. LangGraph suits teams that want explicit graphs on any cloud. Hosting and your team’s language matter more than features.
What replaces Semantic Kernel and AutoGen?
Microsoft folded both into Microsoft Agent Framework, announced on 1 Oct 2025 and generally available as 1.0 on 3 Apr 2026. The AutoGen repository is archived and points to Agent Framework as its successor, and the Semantic Kernel README says it is now Microsoft Agent Framework. Start new projects on Agent Framework.
Is LangChain the same as LangGraph?
They are separate libraries from the same company. LangGraph is the graph runtime with checkpointers and interrupts; LangChain 1.0 provides create_agent and middleware, such as human approval of tool calls and PII redaction, and runs on the LangGraph runtime. Legacy LangChain code moved to langchain-classic.
Can one platform run agents built with different frameworks?
Yes. Foundry hosted agents and AgentCore Runtime both accept several frameworks, and A2A lets agents built on different frameworks call each other. Standardise identity, MCP tools, A2A and OpenTelemetry at the edges, and let each team pick the framework that fits its agent.
More in Platforms and frameworks
Sources
Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.
- Introducing Microsoft Agent Framework
- Microsoft Agent Framework version 1.0
- Agent Framework observability
- AutoGen repository (previous framework, archived)
- Semantic Kernel repository (Semantic Kernel is now Microsoft Agent Framework)
- LangChain and LangGraph 1.0
- LangGraph durable execution
- Introducing Strands Agents 1.0
- Agent Development Kit: easy to build multi-agent applications
- Agent Development Kit
- adk-python v2.0.0 release
- OpenAI Agents SDK
- Building agents with the Claude Agent SDK
- CrewAI introduction
- CrewAI AMP introduction
- Announcing Workflows 1.0
- What is Microsoft Foundry Agent Service?
- Harnesses in Copilot Studio
- What is Amazon Bedrock AgentCore?
- Amazon Bedrock Agents Classic maintenance mode
- Top announcements of the AWS Summit in New York 2026
- Gemini Enterprise Agent Platform release notes
- Claude Platform release notes
- OpenAI API changelog
- Deprecations
- Announcing the OpenAI Agents SDK integration
Planning an AI program?
Talk to the team that builds it and runs it after launch.
- Built to SOC 2, HIPAA and GDPR standards
- EU AI Act-aligned delivery
- Microsoft Solutions Partner