AI Agent Frameworks Compared: Claude SDK, OpenAI, LangGraph

Pick a vendor SDK, such as the Claude Agent SDK or OpenAI Agents SDK, if you want a working agent fast on that vendor. Pick LangGraph if you need explicit control over long, stateful workflows. Pick Pydantic AI, Google ADK, Microsoft Agent Framework or CrewAI when their language or ecosystem fits your team.

A framework is the code that runs your agent's loop, tools and memory. It is separate from the model. To compare models, see the current leaderboard, and our model vs agent vs harness guide explains why the wrapper affects results.

Disclosure: eSEOspace is a member of the Claude Partner Network.

At a glance

Details below come from each project's documentation, read on 2 November 2026.

Do you need a framework at all?

Often not at first. Anthropic's engineering guide advises starting simple and adding complexity only when needed. A model API and a short tool loop can go a long way.

A framework earns its place when you need durable state, approvals, tracing or several agents working together. If you are still deciding whether to write code at all, read build vs buy first.

Claude Agent SDK

The Claude Agent SDK gives you the same tools, loop and context management that power Claude Code, in Python and TypeScript. That means reading files, running commands and editing code work from day one. It also includes hooks, subagents and permissions.

Its strength is how much comes built in. Its limits are real too. It is built for Claude models and runs the Claude Code binary, so you depend on that release cycle.

Anthropic also says third-party developers cannot offer claude.ai login or rate limits in their products without prior approval. So plan on API billing if you ship a product, under Anthropic's Commercial Terms.

For more, see our guide to the Claude Agent SDK for business.

OpenAI Agents SDK

The OpenAI Agents SDK describes itself as a lightweight package with very few abstractions. Its main building blocks are agents, handoffs between agents, and guardrails. It also has built-in tracing, sessions, sandbox agents and MCP support.

Its docs include third-party adapters such as LiteLLM and Any-LLM, so you are not limited to one vendor in principle. Our take: expect the best experience on OpenAI's own models, and test any alternative you plan to use.

It also has a clear human-in-the-loop flow, which pauses a run until someone approves a tool call. See our OpenAI Agents SDK tutorial for a hands-on start.

LangGraph

LangGraph is a low-level orchestration framework and runtime for long-running, stateful agents. You draw your workflow as a graph, with explicit steps and branches. It focuses on durable execution, streaming, human-in-the-loop and persistence.

Durable execution means a run can survive failure and pick up where it left off. Interrupts can pause a graph at any point, save state and wait for a person.

The tradeoff is effort. LangChain's own docs call it very low-level, and they point users wanting something quicker to its higher-level options.

Tracing usually runs through LangSmith, whose paid plan listed $39 per seat per month on 2 November 2026. Our LangGraph tutorial shows what the code looks like. If you run it in production, budget for that tracing cost as well.

Google ADK

Google's Agent Development Kit is an open-source framework available in Python, TypeScript, Go, Java and Kotlin. That language range stands out if your backend is not Python.

It suits teams already invested in Google Cloud. If you use it mainly for Gemini, watch Gemini's price schedule, since listed rates on some models rise from December 2026. We reviewed only its overview page for this guide, so treat it as a candidate to test.

Microsoft Agent Framework

Microsoft Agent Framework combines AutoGen's agent abstractions with Semantic Kernel's enterprise features. It adds graph-based workflows for multi-agent orchestration.

It supports several providers, including Microsoft Foundry, Anthropic, Azure OpenAI, OpenAI and Ollama. It also offers a harness agent for long tasks, with planning, memory and context compaction. The Go version is in public preview and lacks some features.

If your company runs on Microsoft and .NET, this is the natural place to look. If you had AutoGen or Semantic Kernel projects, check how they map before you plan a migration.

CrewAI

CrewAI is an open-source framework built around two ideas. Crews are teams of autonomous agents, while Flows are structured, event-driven workflows that control state and execution.

The team metaphor is easy to explain to non-engineers, which helps early prototypes. CrewAI also has a commercial platform, so ask what stays open source. Multi-agent setups also cost more in tokens, so measure this before you scale.

Pydantic AI

Pydantic AI calls itself a typed, extensible agent loop for Python. You swap models by changing a string, and it can fall back to another model on failure. Its team also offers Pydantic Evals for grading outputs and tool paths, and Pydantic Graph for typed control flow.

This is a good fit for Python teams who care about type checking and testing. Note the vendor claims come from its own site, so verify them in your trial.

Decision factors that matter

Model flexibility

If you might change vendors, prefer a framework with multi-provider support. Vendor SDKs are often tuned to their own models. Compare models on the coding and knowledge work boards, and expect positions to shift.

State and long-running work

Ask what happens if the process crashes at step 40 of 50. If that would be a disaster, look for durable execution and checkpoints. If tasks finish in a minute, you may not need them.

Human approval

Approval should be a built-in feature, not something you add later. Check how the framework pauses, stores state and resumes. Our human-in-the-loop guide helps you decide what needs approval.

Tools and MCP

Most options now support MCP, an open standard for connecting agents to tools. Our comparison of MCP frameworks and SDKs has details. Keep your tools in MCP servers if you want to change frameworks later.

Testing and tracing

You need to see each step an agent takes, and you need repeatable tests. Check what comes included and what costs extra. Use our agent scorecard on a small prototype.

Team skills

The best framework is usually the one your developers can maintain. A low-level graph is no use if nobody can debug it. Ask who will own this code in two years.

How to choose in one week

1. Write down one real workflow, including its approval points.
2. Shortlist two frameworks based on the factors above.
3. Build the same small prototype in each, with two or three tools.
4. Run it on 10 real tasks, and record failures and developer time.
5. Pick the one that was easier to debug, not just faster to demo.

Keep your prompts, tests and tools separate from the framework where you can. That lets you change your mind later at a lower cost.

Which should you choose

Choose the Claude Agent SDK if your agent works with files and code, and you are happy to commit to Claude. Choose the OpenAI Agents SDK if you want a light starting point in that ecosystem.

Choose LangGraph if the workflow is long, has branches and must survive failures. Choose Microsoft Agent Framework or Google ADK if your company's stack points to them.

Choose Pydantic AI if your Python team values typing and wants easy model switching, and CrewAI if a team-of-agents design fits your problem. Check the leaderboard methodology before you let benchmark results drive a framework choice, since those results mostly compare models.

Frequently asked questions