There are now dozens of AI agent frameworks, SDKs and managed platforms, several have been renamed, merged or archived in the last year, and most lists online mix them up. This hub is a dated directory of the notable options as of November 2026: open-source frameworks and SDKs, managed agent platforms from the cloud and model providers, and the two interoperability protocols they share. For each entry it gives the owner, languages, licence, current status and the official source, then shows which frameworks run on which platforms. It is written for engineers choosing a stack, architects mapping what is already in use and security teams building an inventory. Entries are listed alphabetically within each table, and order is never a ranking. Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation and official repositories; it is not sponsored or endorsed by any provider.
Key takeaways
- Three layers, not one market. Open-source frameworks and SDKs give you the agent loop in code, managed platforms run and govern agents for you, and protocols (MCP and A2A) connect agents to tools and to each other.
- Microsoft consolidated. The AutoGen repository is archived and no longer maintained, and the Semantic Kernel repository now points to Microsoft Agent Framework, which its README describes as available at version 1.0 as a production-ready release.
- Licences are mostly permissive, with exceptions. Most repositories checked use MIT or Apache-2.0; Mastra uses a dual model where code in ee/ directories is under the Mastra Enterprise License.
- Platforms host many frameworks. Amazon Bedrock AgentCore names LangChain, OpenAI Agents SDK, Claude Agent SDK and Strands; Microsoft Foundry hosted agents name Agent Framework and LangGraph.
- Status changes matter for maintenance and security. Archived projects receive no security fixes, and Amazon Bedrock Agents Classic is closed to new customers, so the inventory should record status, not just names.
What is an AI agent framework or platform?
An agent framework or SDK is a code library that runs the loop in which a model decides on a step, calls a tool, reads the result and continues until the task is done. It usually adds tool definitions, memory, multi-agent patterns such as handoff or group chat, and tracing hooks. You install it in your own application and run it on your own infrastructure.
A managed agent platform is a cloud service that hosts agents and supplies the surrounding parts: runtime and sandboxes, identity and access, tool gateways, memory, evaluation and observability. Several platforms run agents written in third-party frameworks as well as their own.
A protocol is a shared specification. The Model Context Protocol (MCP) describes itself as an open-source standard for connecting AI applications to external systems such as data sources, tools and workflows. A2A describes itself as an open protocol for communication and interoperability between opaque agentic applications.
The trade-offs between building on a framework and buying a platform are covered in Agent Frameworks vs Managed Agent Platforms, and a side-by-side of platform features, models and prices is in AI Agent Platforms Compared. This hub does not repeat them; it is the index.
Open-source agent frameworks and SDKs
As of 15 November 2026. Licence is taken from each repository's LICENSE file; language and status from the repository README or official docs.
| Framework | Owner | Languages | Licence (repository) | Status, November 2026 |
|---|---|---|---|---|
| AG2 | AG2 community | Python | Apache-2.0 | Active; v1.0 made the protocol-driven framework the top-level ag2 package, and the classic autogen namespace moved to the ag2-classic repository |
| AutoGen | Microsoft | Python | MIT for code (LICENSE-CODE); CC BY 4.0 for documentation | Archived; README states no bug fixes, security fixes, features or support; new users directed to Microsoft Agent Framework |
| Claude Agent SDK | Anthropic | Python, TypeScript | MIT (Python repository) | Active |
| CrewAI | CrewAI Inc | Python (3.10 to below 3.14) | MIT | Active; open-source framework plus a commercial enterprise platform |
| Google Agent Development Kit (ADK) | Python, with Java, Kotlin and Go versions | Apache-2.0 | Active; Python README titled ADK 2.0 | |
| Haystack | deepset | Python | Apache-2.0 | Active; pipelines and agent workflows |
| LangChain | LangChain Inc | Python, with LangChain.js for JavaScript and TypeScript | MIT | Active |
| LangGraph | LangChain Inc | Python, with LangGraph.js for JavaScript and TypeScript | MIT | Active; low-level orchestration for stateful agents |
| LlamaIndex | LlamaIndex | Python | MIT | Available; README states the company's primary focus has shifted to LlamaParse document parsing, with the framework still offered as an open toolkit |
| Mastra | Mastra | TypeScript | Apache-2.0 for the core; Mastra Enterprise License for ee/ directories | Active |
| Microsoft Agent Framework | Microsoft | Python, .NET, Go (separate repository) | MIT | Version 1.0, described as production-ready with stable APIs; successor to AutoGen and Semantic Kernel |
| OpenAI Agents SDK | OpenAI | Python, JavaScript and TypeScript | MIT | Active; supports the Responses and Chat Completions APIs and, per its README, 100+ other LLMs |
| Pydantic AI | Pydantic | Python | MIT | Active |
| Semantic Kernel | Microsoft | Python, .NET, Java | MIT | README states Semantic Kernel is now Microsoft Agent Framework and links a migration guide |
| smolagents | Hugging Face | Python | Apache-2.0 | Active; lightweight, LLM-agnostic library |
| Strands Agents | Amazon Web Services | Python, TypeScript | Apache-2.0 | Active; runs in your process with no hosted control plane |
| Vercel AI SDK | Vercel | TypeScript | Apache-2.0 | Active; provider-agnostic, uses the Vercel AI Gateway by default |
What each framework is built around
- AG2: agents that interact with each other to solve tasks, aimed at development and research of agentic AI.
- AutoGen: multi-agent applications that act autonomously or alongside humans; kept for reference only.
- Claude Agent SDK: the agent harness behind Claude Code as a library, with built-in tools, hooks, subagents, MCP and permissions controlling which tools run without approval.
- CrewAI: high-level abstractions and low-level APIs for multi-agent workflows, organised as role-based crews and flows.
- Google ADK: a code-first framework for building, evaluating and deploying agents.
- Haystack: modular pipelines and agent workflows with explicit control over retrieval, routing, memory and generation.
- LangChain: quick-start agents with any model provider, described by LangChain Inc as part of its agent engineering platform.
- LangGraph: graph-based, stateful orchestration with low-level control.
- LlamaIndex: orchestration tools for retrieval-augmented generation and agent applications, often used with LlamaParse.
- Mastra: agents, workflows and model routing across 40+ providers in a TypeScript stack that integrates with React, Next.js and Node.
- Microsoft Agent Framework: agents and multi-agent workflows with sequential, concurrent, handoff and group collaboration patterns, durability and human-in-the-loop control.
- OpenAI Agents SDK: agents with instructions, tools, guardrails and handoffs, plus sandbox agents, realtime agents and voice pipelines.
- Pydantic AI: typed agents in Python, including realtime voice, image generation and embeddings.
- Semantic Kernel: a model-agnostic SDK for agents and multi-agent systems, now superseded.
- smolagents: agents in a few lines of code, including code agents.
- Strands Agents: an SDK for teams that would otherwise write their own agent loop, with CLI, MCP server and harness components.
- Vercel AI SDK: a unified API across model providers for AI applications and agents in Next.js, React, Svelte, Vue, Angular and Node.js.
For framework-specific depth see LangChain and LangGraph, and for the order in which these projects appeared and changed see the AI Agent Frameworks Timeline.
Managed agent platforms
As of 15 November 2026, from each provider's product or documentation page. Pricing is covered in AI Agent Platforms Compared.
| Platform | Owner | What the provider says it is | Status, November 2026 |
|---|---|---|---|
| Amazon Bedrock AgentCore | Amazon Web Services | Services to deploy and operate agents: Harness, Gateway (authentication and access control for MCP servers, APIs and Lambda functions), Evaluations, tracing and security policies | Active |
| Amazon Bedrock Agents | Amazon Web Services | Agents that use model reasoning, APIs and data to complete tasks, with multi-agent collaboration, retrieval, memory, code interpretation and Guardrails | Agents Classic no longer available to new customers; AWS points to AgentCore |
| Claude Managed Agents | Anthropic | A pre-built, configurable agent harness running Claude on Anthropic-managed infrastructure, with stateful sessions and built-in tools | Beta (managed-agents-2026-04-01 header) |
| CrewAI enterprise platform | CrewAI Inc | Enterprise agent build and runtime: visual builder exportable to Python, control plane with tracing, cost accounting, RBAC, audit trails and approval gates | Active |
| Gemini Enterprise Agent Platform (page titled "Gemini platform") | Google Cloud | Platform for building, scaling, governing and optimising agents, including Agent Studio, ADK, Antigravity and Model Garden with 200+ models | Active |
| LangSmith | LangChain Inc | Agent engineering platform: observability, evaluation, Agent Server, sandboxes, Fleet and an LLM Gateway | Active; SaaS, BYOC or self-hosted |
| Microsoft Foundry Agent Service | Microsoft | Fully managed platform for building, deploying and scaling agents: prompt agents, workflow agents (preview) and hosted agents (preview) | Active |
| OpenAI AgentKit | OpenAI | Agent Builder visual canvas, Connector Registry, ChatKit, Evals and Guardrails | Launched 6 October 2025 with Agent Builder in beta and ChatKit and Evals generally available |
Which frameworks run on which platforms?
Only pairings that the platform's own page names are listed. "Your own framework" means the page says any framework can be used.
| Platform | Frameworks named by the provider | Model choice stated |
|---|---|---|
| Amazon Bedrock AgentCore | Claude Agent SDK, LangChain, OpenAI Agents SDK, Strands SDK, your own framework | Any model |
| Claude Managed Agents | Anthropic's own harness; agents defined by model, system prompt, tools, MCP servers and skills | Claude models on the Claude API, Claude Platform on AWS and Google Cloud |
| CrewAI enterprise platform | CrewAI, with evaluation integrations for Arize, Datadog, Galileo and Patronus | Testing across multiple LLMs |
| Gemini Enterprise Agent Platform | Agent Development Kit (ADK) | Google, third-party and open models through Model Garden |
| LangSmith | Deep Agents, LangChain, LangGraph | LLM Gateway controls model calls |
| Microsoft Foundry Agent Service | Agent Framework, LangGraph, your own framework (hosted agents) | Foundry model catalogue, swappable without code changes |
| OpenAI AgentKit | Agent Builder workflows; Evals support third-party models | OpenAI models through the API |
Amazon Bedrock Agents is omitted from this table because new customers are directed to AgentCore.
Interoperability protocols
| Protocol | Purpose | Governance and version stated on the official site | SDKs |
|---|---|---|---|
| A2A | Communication and interoperability between opaque agentic applications | Linux Foundation neutral governance, part of the Agentic AI Foundation; A2A v1.0 | Six official SDKs: C#/.NET, Go, Java, JavaScript, Python, Rust |
| Model Context Protocol (MCP) | Connecting AI applications to data sources, tools and workflows | Documentation versioned 2026-07-28 at the time of checking | Client and server build guides on the official site |
Most frameworks and platforms in this hub can act as MCP clients: AgentCore Gateway, Foundry Agent Service, Claude Managed Agents and the Strands components all reference MCP on their own pages. For the protocol's security model see the MCP Security Guide and the A2A Protocol Security Guide.
What changed in 2025 and 2026?
| Date or version | Change | Source |
|---|---|---|
| 6 October 2025 | OpenAI launches AgentKit; Agent Builder beta, ChatKit and Evals generally available | OpenAI announcement |
| 2026 (beta header dated 2026-04-01) | Claude Managed Agents in beta, enabled by default for all API accounts | Claude docs |
| Version 1.0 | Microsoft Agent Framework production-ready release; Semantic Kernel README points to it | Microsoft repositories |
| Archived | AutoGen repository archived with no further fixes | AutoGen README |
| AG2 v1.0 | Protocol-driven framework becomes the ag2 package; classic API moves to ag2-classic | AG2 README |
| ADK 2.0 | Google's Python ADK README carries the 2.0 title | ADK README |
| A2A v1.0 | A2A at v1.0 under the Linux Foundation and Agentic AI Foundation | A2A site |
| Agents Classic closed | Amazon Bedrock Agents Classic no longer available to new customers | AWS product page |
Exact dates for several of these changes are not stated on the pages checked, so they are shown by version or status. The AI Agent Frameworks Timeline tracks dated events.
Which migration paths do providers publish?
Where a project has been superseded or closed, the provider's own page names the destination. These are the paths stated as of 15 November 2026.
| From | To | What the provider states |
|---|---|---|
| Amazon Bedrock Agents Classic | Amazon Bedrock AgentCore | Agents Classic is no longer available to new customers; AWS directs users to AgentCore for similar capabilities |
| AutoGen | Microsoft Agent Framework | New users should start with Agent Framework; existing users are encouraged to migrate using the AutoGen to Agent Framework migration guide on Microsoft Learn |
| AutoGen classic API (ConversableAgent, GroupChat, import autogen) | AG2 Classic | The AG2 README states these now live in the ag2-classic repository, and the ag2 package no longer ships the autogen import name |
| Semantic Kernel | Microsoft Agent Framework | The Semantic Kernel README calls Agent Framework its enterprise-ready successor and links a migration guide |
A migration changes more than code. Agent identities, tool credentials, logging and approval steps are often tied to the old runtime, so plan the move as an identity and access change as well as a dependency upgrade.
How to choose: factual criteria
These are criteria to check against each option's own documentation, not a ranking.
| Criterion | What to check | Where it shows in this hub |
|---|---|---|
| Language | Does the framework support your team's main language (Python, TypeScript, .NET, Java, Go)? | Frameworks table |
| Maintenance status | Is the project active, superseded or archived, and does it still receive security fixes? | Status column |
| Licence | Is the code permissive (MIT, Apache-2.0) or are some parts under a commercial licence? | Licence column |
| Hosting | Will agents run in your process, on your cloud, or on a provider-managed runtime? | Platforms and pairing tables |
| Model portability | Can you change model provider without rewriting agent code? | Pairing table, model choice column |
| Identity and access | Does each agent get its own identity, and how are tool credentials issued? | Platform descriptions; see security section |
| Protocol support | Does it act as an MCP client or server, and does it speak A2A? | Protocols section |
Small teams often start with one framework in one language and add a platform when they need hosted runtime, evaluation or governance. Large organisations usually find several frameworks already in use and need an inventory first, then a short approved list and a platform that hosts more than one of them.
Where identity and security fit
Every framework and platform here runs agents that hold credentials: model API keys, tool tokens, cloud roles and OAuth grants. That makes agents non-human identities to inventory and govern. The platform pages state specific controls, for example per-agent Microsoft Entra identities and Azure RBAC in Foundry Agent Service, Gateway authentication in AgentCore, and RBAC and audit trails in the CrewAI enterprise platform. Archived projects such as AutoGen receive no security fixes, which is itself a risk to record.
Security guides per framework or platform: AWS Bedrock AgentCore Security Guide, Claude Agent SDK Security Guide, CrewAI security guide, Google Agent Development Kit Security Guide, LangChain and LangGraph Security Guide, LlamaIndex Security Guide, Microsoft Agent Framework and AutoGen security guide, Microsoft Foundry Agent Service Security Guide, Securing OpenAI Agents SDK and Responses API Guide, Pydantic AI Security Guide, Semantic Kernel Security Guide, AWS Strands Agents SDK Security Guide. The cross-framework view is in the Securing Agent Frameworks Guide.
Practitioner checklist
- Inventory every agent framework and platform in use, including in prototypes and notebooks, with owner, version and status.
- Flag archived or superseded projects (AutoGen, Semantic Kernel) and plan migration where the provider publishes a guide.
- Record each repository's licence, and check for commercially licensed directories such as Mastra's ee/ code.
- For each platform, record which frameworks it hosts and which models it can call.
- List the credentials each agent holds and give agents their own identities where the platform supports it.
- List MCP servers and A2A connections per agent, with their authentication method.
- Re-check status and licence columns against official sources at least quarterly, because names and support change.
Standards and references
- Model Context Protocol: What is MCP?
- A2A Protocol: Official site
- Amazon Web Services: Amazon Bedrock AgentCore
- Microsoft Learn: Foundry Agent Service overview
- Google Cloud: Gemini Enterprise Agent Platform
- OpenAI: Introducing AgentKit
- Anthropic: Claude Managed Agents overview
- LangChain: LangSmith and open-source frameworks
Related NHI Mgmt Group resources: AI Agent Platforms Compared · Agent Frameworks vs Managed Agent Platforms · Securing Agent Frameworks Guide · AI Coding Tools Hub