The Short Answer
The open-source AI agent frameworks enterprises shortlist are LangGraph, CrewAI, Microsoft Agent Framework, the OpenAI Agents SDK and Google ADK, all permissively licensed, and LangGraph is a common pick for long-running stateful workflows. A framework is a library, not where agents run: ibl.ai is the platform layer above it, with SSO, RBAC, audit and any model, where you own all the code and the data.
Every fact about each framework below comes from its own repository, package registry or documentation, checked on 18 November 2026. Versions move weekly, so read them as a snapshot.
Which open-source AI agent frameworks do enterprises actually shortlist in 2026?
Enterprises shortlist six agent frameworks in 2026: LangGraph, CrewAI, Microsoft Agent Framework, the OpenAI Agents SDK, the Claude Agent SDK and Google's Agent Development Kit (ADK). Five carry an OSI-approved licence; the sixth is a special case.
| Framework | Maintainer | Licence | Latest release (18 Nov 2026) | Commercial platform layer |
|---|---|---|---|---|
| LangGraph | LangChain | MIT | 1.2.22 (14 Nov) | LangSmith Deployment |
| CrewAI | CrewAI | MIT | 1.15.42 (18 Nov) | CrewAI AMP Suite |
| Microsoft Agent Framework | Microsoft | MIT | Python 1.26.0 (12 Nov); 1.0 on 3 Apr 2026 | Microsoft Foundry hosted agents |
| OpenAI Agents SDK | OpenAI | MIT | 0.25.0 (18 Nov), pre-1.0 | OpenAI tracing dashboard |
| Claude Agent SDK | Anthropic | MIT file; use under Anthropic Commercial Terms | 0.2.184 (18 Nov) | Claude Managed Agents |
| Google ADK | Apache 2.0 | 2.14.1 (12 Nov) | Agent Runtime on Gemini Enterprise Agent Platform |
Releases are each project's Python package on PyPI. Microsoft's licence and version come from its PyPI package and its 1.0 announcement, which calls the framework "100% open source".
For Microsoft Agent Framework we cite the MIT-licensed package on PyPI for its licence and version.
How do LangGraph, CrewAI and the other agent frameworks differ?
The six agent frameworks differ mainly in how they model state and who they assume you are. Each project's own README describes its centre of gravity.
LangGraph calls itself "a low-level orchestration framework" for long-running, stateful agents, with durable execution, human-in-the-loop interrupts and short- and long-term memory.
CrewAI orchestrates "role-playing, autonomous AI agents" as crews, with memory, knowledge, checkpointing and MCP/A2A support.
Microsoft Agent Framework merged Semantic Kernel and AutoGen; its 1.0 covers .NET and Python, a graph workflow engine with checkpointing, YAML agent definitions and MCP.
OpenAI Agents SDK is "provider-agnostic", supporting OpenAI's APIs "as well as 100+ other LLMs", with handoffs, guardrails, sessions, tracing and sandbox agents.
Claude Agent SDK gives you "Claude Code as a library": built-in file and shell tools, hooks, subagents, permissions, sessions and MCP.
Google ADK is code-first and, in its own words, "optimized for Gemini" yet "model-agnostic, deployment-agnostic".
For a deeper look at the Microsoft consolidation and what it means for where agents execute, see the framework war is about who owns the agent runtime.
Are all of these AI agent frameworks really open source and self-hostable?
Five of the six agent frameworks are open source under OSI-approved licences and run anywhere Python runs: LangGraph, CrewAI, Microsoft Agent Framework and the OpenAI Agents SDK under MIT, Google ADK under Apache 2.0. You can run every one of them on your own infrastructure.
The Claude Agent SDK is the exception worth reading twice.
Its repository ships an MIT licence file, but its README states that use "is governed by Anthropic's Commercial Terms of Service", and it runs the bundled Claude Code binary. It is public source built around one vendor's models.
Self-hostable library code is not the same as a self-hosted production platform. The framework's own hosted layer often carries different terms, which is where the next question starts.
What does an open-source agent framework leave to you in production?
An open-source agent framework leaves the platform layer to you: who may use which agent, under which identity, with which model, and who changed it last. Each vendor's README points to a commercial product for exactly that layer.
CrewAI's README says organizations "that need a commercial control plane" get managed deployment, observability, governance and security from CrewAI AMP, with on-premise and cloud options. LangGraph's README points to LangSmith Deployment for production.
LangSmith documents four hosting options. Its cloud option "Requires a Plus plan or above", and the self-hosted control plane "Requires the Enterprise plan with Agent Server hosting enabled". A standalone Agent Server runs without a control plane but needs a LangSmith license.
| Production need | Open-source framework alone | Platform layer |
|---|---|---|
| SSO and role-based access | You build it | Roles, policies, groups and teams per organization |
| Audit trail | Traces of runs; who changed the agent is yours to record | Per-agent log of who changed what and when |
| Multi-tenant hosting | You build it | Organization isolation with per-tenant configuration |
| Model choice per agent | In code, per deployment | An admin setting on each agent |
| Who owns that layer | Your engineers, or the framework vendor's commercial plan | With ibl.ai, you: source code and data on your infrastructure |
None of this makes the frameworks weaker. It is a division of labour: a library decides how an agent thinks, a platform decides who may use it and on what terms.
Where do open-source agent frameworks run in production with ibl.ai?
ibl.ai is the platform layer, and the frameworks fit inside it rather than compete with it. ibl.ai's own backend documentation lists a LangChain runtime layer that resolves registered MCP connections into LangChain tools at inference time.
On the ibl.ai platform each agent gets role-based access built from roles and policies, an audit log of who changed its model or tools, per-organization isolation, and an LLM picked per agent from any provider.
The link to whatever framework your team picks is MCP. An ibl.ai agent can attach any MCP server URL, and the hosted iblai-agent-chat MCP server lets any MCP client, including Claude Agent SDK code, hold a conversation with a deployed ibl.ai agent.
The Agentic OS is the flagship that every agent runs on, and Agentic Ontology serves your business terms to any agent as MCP servers from your own server. For agent runtimes such as OpenClaw, see the OpenClaw service.
Which parts of ibl.ai are open source, and which are licensed?
ibl.ai is part open source and part licensed, and the line is worth stating exactly. The application and toolkit repositories are public on GitHub under MIT: iblai/os, iblai/base, iblai/lms, iblai/infra-cli, iblai/ontology and the .iblai agent standard.
iblai/base (renamed from iblai/vibe on 18 November 2026) scaffolds a Next.js app on ibl.ai with SSO, AI chat and analytics, and ships the platform skills. The full catalogue is on the developer hub.
The backend and infrastructure are not MIT. They come under a perpetual licence, and the complete source code is transferred in the Codebase Transfer engagement, priced as a "Custom quote" on the pricing page. The full code license explains the terms.
How should an enterprise choose an AI agent framework and a platform?
Choose the agent framework and the platform as two separate decisions, because they lock you into different things. Four questions settle both.
What language and state model does your team need? Stateful graphs point to LangGraph or Microsoft Agent Framework; role-based crews to CrewAI; a light loop to the OpenAI Agents SDK; Google Cloud estates to ADK.
Do you need any model, or one vendor's? OpenAI's SDK and Google ADK say in their READMEs that they work across model providers; the Claude Agent SDK is built for Claude.
Who will run SSO, RBAC and audit? Your engineers, the framework vendor's commercial plan, or a platform you deploy.
Must you own that platform layer? If the answer is yes, you need its source code, not an account on it.
For the platform-side comparison, including Copilot Studio, Bedrock AgentCore and watsonx Orchestrate, read which platform should you use to build and deploy AI agents. For a head-to-head on one framework, see ibl.ai vs LangGraph.
With ibl.ai you own all the code and the data: it is model-agnostic, carries no per-seat pricing, and you can deploy anywhere, from your own cloud to a fully air-gapped network.
ibl.ai is family-owned and operated from New York, NY, and our team stays on for updates and new agents after deployment.