AI Agent Frameworks Compared (2026): LangGraph, CrewAI, ADK and 7 More

I get asked which AI agent framework to use about once a week, usually by someone who has already built the same prototype in three of them and is now worried about picking wrong. The honest answer in October 2026 is that the frameworks have converged on the same ideas (tools, a loop, state, handoffs, a checkpoint you can resume from) and differ mostly in who maintains them, which model vendor they lean towards, and how much of the control flow you are allowed to see. This post compares ten AI agent frameworks as they stand today: LangGraph, CrewAI, the OpenAI Agents SDK, the Claude Agent SDK, Google ADK, Microsoft Agent Framework (the successor to AutoGen and Semantic Kernel), Pydantic AI, Mastra and LlamaIndex. Versions, stars and licences come from the GitHub API on 20 October 2026; I ran the queries myself and they will drift by next week.

What changed in agent frameworks between 2025 and 2026?

Three things happened in roughly twelve months, and together they make most 2025 comparison posts wrong.

First, the big Python frameworks reached 1.0 and then kept going. LangGraph 1.0.0 was tagged on 17 October 2025 and the library is at 1.2.16 today. CrewAI 1.0.0 followed on 20 October 2025 and is now at 1.15.28. Pydantic AI went to 1.0 in September 2025 and then to 2.0 on 23 June 2026; its changelog lists the breaking changes, including OpenAI model names now routing through the Responses API by default. Google's ADK 2.0 went GA for Python on 19 May 2026, for Go on 30 June and for TypeScript on 21 August, and swapped the hierarchical agent executor for a graph-based workflow runtime.

Second, Microsoft consolidated. Microsoft Agent Framework 1.0 shipped on 3 April 2026 for .NET and Python, merging Semantic Kernel and AutoGen. The AutoGen README now carries a caution banner saying the project is archived, that Microsoft is "not providing bug fixes, security fixes, new features, or support", and that new users should start with Agent Framework. The Semantic Kernel README calls Agent Framework its "enterprise-ready successor".

Third, the model vendors shipped their own harnesses. OpenAI's Agents SDK added Sandbox Agents in v0.14.0 on 15 April 2026: a persistent, isolated workspace with manifests, snapshots and resume. Anthropic renamed the Claude Code SDK to the Claude Agent SDK, which exposes the same agent loop, file tools and permission model that Claude Code runs on.

The convergence is the part worth noticing. LangGraph has had checkpoints and interrupt() for a while; ADK 2.0 and Agent Framework now both advertise graph workflows with human-in-the-loop pause and resume. When four teams independently arrive at "a state machine with durable checkpoints", that is probably what an agent framework is.

The comparison table: versions, licences and stars

Numbers below are from gh api repos/OWNER/REPO and releases/latest, run on 20 October 2026. Stars measure attention over a project's whole life, not current health, which the AutoGen row demonstrates nicely.

Framework	Latest release (date)	Licence	GitHub stars	Languages	Best at	Source
LangGraph	1.2.16 (16 Oct 2026)	MIT	43,540	Python, JS	Explicit graphs, checkpoints, interrupt()	repo
CrewAI	1.15.28 (14 Oct 2026)	MIT	59,890	Python	Role-based crews plus event-driven Flows	repo
OpenAI Agents SDK	v0.24.0 (15 Oct 2026)	MIT	30,152	Python, JS	Handoffs, guardrails, sessions, sandbox agents	repo
Claude Agent SDK	v0.2.168 (15 Oct 2026)	Anthropic Commercial Terms	8,302	Python, TS	Claude Code's loop, file and shell tools	repo
Google ADK	v2.12.0 (13 Oct 2026)	Apache-2.0	21,915	Python, Go, TS, Java	Graph workflow runtime, Vertex AI deployment	repo
Microsoft Agent Framework	python-1.22.0 (15 Oct 2026)	MIT	14,217	Python, .NET, Go	Azure and Foundry shops, SK and AutoGen migrations	repo
Pydantic AI	v2.57.0 (16 Oct 2026)	MIT	20,744	Python	Typed outputs, dependency injection	repo
Mastra	@mastra/core 1.78.0 (16 Oct 2026)	Apache-2.0 plus ee/ licence	28,927	TypeScript	TS-native agents, workflows, memory	repo
LlamaIndex	v0.14.25 (21 Sep 2026)	MIT	52,572	Python, TS	Retrieval-heavy agents, Workflows	repo
AutoGen	python-v0.7.5 (30 Sep 2025)	MIT code, CC-BY-4.0 docs	61,493	Python, .NET	Nothing new: archived, migrate	repo

Two licence details trip people up. The Claude Agent SDK Python repo has an MIT LICENSE file, but its README says use of the SDK "is governed by Anthropic's Commercial Terms of Service", and the TypeScript repo's licence file says the same thing outright. Mastra's licence file puts everything under Apache-2.0 except directories named ee/, which carry a separate enterprise licence. Neither is a problem for most teams. Both are worth a line in your dependency review.

LangGraph vs CrewAI vs OpenAI Agents SDK: how they differ

The useful axis is not features but how much control flow the framework hides from you.

LangGraph hides almost nothing. You declare nodes, edges and a state schema, compile a graph, and pass a checkpointer. Every step writes a checkpoint keyed by thread_id, which is what makes pause, resume, time travel and human approval cheap to add. The cost is verbosity: a simple tool-calling loop is thirty lines where other frameworks need five. If you need an approval gate before a destructive tool, this is the framework where that is a first-class primitive, and I wrote a step-by-step LangGraph human-in-the-loop guide with Postgres persistence to show the whole path.

CrewAI starts at the other end: agents with roles, goals and backstories, grouped into a crew that works through tasks. Its README now positions Flows as the "precise, event-driven control" layer next to crews, which is CrewAI admitting, sensibly, that production work needs deterministic orchestration around the autonomous parts. It is quick to demo and the role metaphor is easy to explain to non-engineers. Debugging a crew that took an unexpected path is harder than debugging an explicit graph.

The OpenAI Agents SDK sits in the middle. The core is small: agents with instructions and tools, handoffs between agents, guardrails that validate input and output, and sessions for conversation history. Its README says it is provider-agnostic, supporting the Responses and Chat Completions APIs "as well as 100+ other LLMs". The April sandbox release is the bigger story: agents that work over a real filesystem and shell, with snapshots and resume. That moves the SDK from "orchestration library" into "harness", and it means you now have to think about where that sandbox runs.

Claude Agent SDK, Google ADK and Microsoft Agent Framework

The Claude Agent SDK is a different kind of thing. It is not an orchestration graph; it is the Claude Code agent loop, packaged. You get the built-in tools (read, write, edit, bash, search), hooks, subagents, MCP servers and a permission system, and you call it from Python or TypeScript. For agents that do work on files and code it is the shortest path to something capable, because the hard parts (context management, tool use, editing) are already tuned. The trade-offs are vendor lock-in to Claude models and the commercial terms noted above.

Google ADK is now the most multi-language option, with first-party repos for Python, Go, TypeScript and Java. ADK 2.0's release notes describe agents, tools and functions running as nodes in a workflow graph, and warn that custom session storage with rigid schemas must handle new event fields, and that broad exception handling in tools hides failures from the retry mechanism. If you run on Google Cloud and want Vertex AI Agent Engine as the deploy target, this is the default choice.

Microsoft Agent Framework is the answer for .NET teams and anyone already on Semantic Kernel or AutoGen. The 1.0 announcement promises stable APIs and long-term support, and Microsoft publishes migration guides from AutoGen and from Semantic Kernel. If you are still running AutoGen in production, that migration is no longer optional; an unpatched dependency in an agent with tool access is exactly the kind of risk I covered in MCP Security in 2026.

Pydantic AI, Mastra and LlamaIndex: the specialists

Pydantic AI is what I reach for when the agent is one component inside a normal Python service. Typed dependencies, validated structured outputs and model-agnostic providers make it feel like FastAPI rather than a framework with opinions about your architecture. The 2.0 jump matters if you pin loosely: a bare install now ships fewer provider extras, so Bedrock, Groq and Mistral need explicit installation.

Mastra is the TypeScript-native option, with agents, workflows, memory and evals in one package. @mastra/core 1.0.0 was tagged on 20 January 2026 and has shipped 78 minor releases since, which tells you both that it is moving fast and that you should pin it. If your product is a Next.js app, writing the agent in the same language is worth more than any feature on this list.

LlamaIndex remains strongest when retrieval is the job. Its event-driven Workflows now live in a separate llama-index-workflows package, and the main repo's release cadence is slower than the others (latest release 21 September). For agentic RAG over messy document sets it is still a sensible default.

How to choose an agent framework for production

Here is the decision as I would actually make it, as a dependency pin you can paste into a project and adjust:

# pyproject.toml: pick ONE orchestration layer, pin it exactly, review upgrades like any other dependency.
[project]
name = "ops-agent"
requires-python = ">=3.11"
dependencies = [
  # Explicit graph, durable checkpoints, human approval gates:
  "langgraph==1.2.16",
  "langgraph-checkpoint-postgres==3.2.0",
  # OR typed single-agent inside an existing service:
  # "pydantic-ai==2.57.0",
  # OR OpenAI-first with handoffs and sandboxes:
  # "openai-agents==0.24.0",
  # OR Google Cloud and Vertex AI Agent Engine:
  # "google-adk==2.12.0",
  # OR Azure or .NET shop, or migrating from AutoGen / Semantic Kernel:
  # "agent-framework==1.22.0",
]

The rules behind that list:

1. If a human must approve side effects, pick something with durable checkpoints. LangGraph, ADK 2.0, Agent Framework and Mastra's suspend and resume all qualify now. An approval that lives in process memory dies on the next deploy.
2. If you are locked to one model vendor anyway, use its SDK. The OpenAI Agents SDK and Claude Agent SDK get new model features first and need less glue.
3. Match the language of the product. TypeScript shops should look at Mastra or the JS ports before forcing a Python service into the stack.
4. Do not choose by stars. The most-starred repo in the table is the one you should not start on.
5. Budget for the parts no framework gives you. Isolation for code execution, identity for the agent, traces you can query, and policy on tool calls. The agent observability and sandbox questions are the same whichever framework you pick.

That last point is where platform discipline still earns its keep. Every framework here will happily run rm -rf if a model asks and a tool allows it. The framework decides the loop; your infrastructure decides the blast radius. (Disclosure: I run PandaStack, a Firecracker microVM cloud for agent code execution, so I have an obvious interest in that layer; the sandbox runtime comparison covers the options neutrally.) For the conceptual background on loops, planners and multi-agent patterns, see what AI agents are and how they are built.

FAQ

What is the best AI agent framework in 2026?

There is no single best one. LangGraph is the strongest general choice when you need explicit control, persistence and human approval; Pydantic AI is the lightest for typed agents inside a Python service; the OpenAI Agents SDK and Claude Agent SDK are best if you are committed to one model vendor. Pick by language, vendor and whether you need durable state.

Is AutoGen dead?

Effectively, yes. The AutoGen README says the project is archived and that Microsoft provides no bug fixes, security fixes or support, and directs new users to Microsoft Agent Framework, which reached 1.0 on 3 April 2026. Existing users should follow Microsoft's AutoGen migration guide.

LangGraph vs CrewAI: which should I use?

Use CrewAI if you want role-based multi-agent collaboration running quickly and can live with less visibility into each step. Use LangGraph if you need to see and control every transition, resume from checkpoints, or insert approval gates. Many teams prototype in CrewAI and rebuild the production path as an explicit graph.

Is the Claude Agent SDK open source?

Not in the usual sense. The Python repository contains an MIT licence file, but both SDK READMEs state that use is governed by Anthropic's Commercial Terms of Service, and the TypeScript repo's licence file says so directly. Treat it as a vendor SDK, not an open-source framework.

If I were starting an agent project this quarter, I would pick LangGraph with the Postgres checkpointer for anything that touches production systems, Pydantic AI for small typed agents embedded in existing services, and I would spend more time on the sandbox, identity and tracing around them than on the framework choice itself. I would also pin every version in that table exactly, because half of them shipped a release in the last week.

Related: AI Agent Sandbox Runtime: The Six Options and Who Uses Them, Agent-to-Agent Protocols in 2026: A2A Reached 1.0, Three ACPs Died or Merged, and MCP Learned to Do Most of the Same Job and Policy as Code for AI Agents in 2026: Cedar at the Gateway, Rego in the Hook, and the 11.2% the Model Still Lets Through.