AI Engineering5 min read
Google ADK, LangGraph, OpenAI, Claude and CrewAI Compared (2026)
Which agent framework to pick in 2026: Google ADK 2.0, LangGraph, OpenAI Agents SDK, Claude Agent SDK and CrewAI compared on control flow, human approval, models, deployment and RAG.
Ajeet Kumar
Most agent projects do not fail because of the framework. They fail on retrieval quality, missing approvals and no evaluation. The framework still matters, though: it decides how much control flow you write yourself, how a run pauses for a person, which models you can use and where the agent runs. This is how we compare the five frameworks we use or evaluate most, as of November 2026.
The short answer
- Google ADK 2.0: you want one framework from local dev UI to evaluation to deployment, especially on Google Cloud, with graph workflows and human input built in.
- LangGraph: you need fine-grained, durable control over state, with checkpoints in your own Postgres and pauses that can last days.
- OpenAI Agents SDK: you want the smallest set of primitives (agents, handoffs, guardrails, tracing) and a quick path to a working agent.
- Claude Agent SDK: the agent works with files, code and a shell, like Claude Code, and you want that harness rather than building it.
- CrewAI: you want a fast prototype of role-based agents working together, and you accept less low-level control.
Side by side
| Google ADK 2.0 | LangGraph | OpenAI Agents SDK | Claude Agent SDK | CrewAI | |
|---|---|---|---|---|---|
| Model | Graph of nodes plus prebuilt sequential, parallel and loop agents | Low-level state graph | Agents with handoffs | Agent loop with built-in tools | Role-based crews and flows |
| Languages | Python, TypeScript, Go, Java, Kotlin | Python, JavaScript | Python, TypeScript | Python, TypeScript | Python |
| Human approval | RequestInput nodes, tool confirmation | Interrupts with checkpoints | Tool approvals that pause a run; RunState resumes it | Permission modes and hooks | Human input on tasks |
| Models | Gemini native; Anthropic, OpenAI, LiteLLM, Ollama, vLLM | Any via integrations | OpenAI first; others via adapters | Claude | Any via LiteLLM |
| Tools and protocols | Function tools, MCP, A2A, OpenAPI | Tools, MCP adapters | Function tools, MCP | MCP, subagents, hooks | Tools, MCP |
| Deployment | Cloud Run, GKE, Agent Runtime, any container | Any container, LangGraph Platform | Any container | Any container | Any container, CrewAI platform |
| Evaluation | adk eval, pytest, conformance tests | LangSmith | Tracing; evals in your stack | Your own | Your own or platform |
The table describes defaults and first-party features. With enough code, any of them can call any model or run anywhere; the question is how much of that code you want to own.
Google ADK 2.0
ADK reached 2.0 general availability for Python in May 2026, Go in June and TypeScript in August. The headline is the workflow runtime: agents, tools and functions are nodes in a graph that can route, fan out, pause for human input and resume. Around it ADK ships the pieces teams usually assemble themselves: a local web UI with traces, a CLI, an evaluation framework, callbacks for guardrails, MCP and A2A support, and deploy commands for Cloud Run, GKE and Agent Runtime on Gemini Enterprise Agent Platform (Vertex AI until April 2026). Gemini is the native model, and Anthropic, OpenAI and local models are documented too.
Watch for: teams upgrading from 1.x must handle breaking changes in the event schema and custom agent classes, and some integrations are not yet compatible with graph workflows.
LangGraph
LangGraph reached 1.0 in October 2025 and remains the most explicit option: you define the state, the nodes and the edges, and checkpointers persist every step, so a run can pause for approval and resume days later, even on another machine. That control is why we use it for long-running, regulated workflows. You pay for it with more code, and observability is best through LangSmith.
OpenAI Agents SDK
The Agents SDK keeps the surface small: agents with instructions and tools, handoffs between agents, guardrails that validate input and output, sessions and built-in tracing. Tools can require approval: the run pauses with an interruption, and RunState lets you save it and resume after a person decides. It is the fastest way to a working multi-agent prototype on OpenAI models and can reach other providers through adapters. Complex branching stays in your own code.
Claude Agent SDK
The Claude Agent SDK packages the agent harness behind Claude Code: an agent loop with tools for files, shell commands and search, subagents, hooks and MCP. It is a strong choice when the agent's job looks like an engineer's job, such as working through a repository, a data folder or a ticket queue, and you want to start from a proven loop.
CrewAI
CrewAI organises work as crews of role-based agents with tasks, plus flows for more structured sequences. It is quick to demo and easy to explain to stakeholders. For systems that must survive failures and pause for approvals, we usually move the orchestration to LangGraph or ADK.
What about RAG?
None of these frameworks is your retrieval system. Retrieval quality comes from parsing, chunking, embeddings, filtering and reranking, and from measuring them. In all five, retrieval is a tool: a function that queries pgvector, Pinecone, Qdrant or Weaviate, or a managed service such as RAG Engine on Google Cloud, which ADK can call with a built-in tool. Pick the framework for orchestration, then build retrieval with LlamaIndex or LangChain components and evaluate it with metrics such as faithfulness and context precision.
How we choose for clients
- Start from the riskiest step. If an action moves money or changes records, pick the framework whose approval and resume story you trust most: LangGraph or ADK 2.0.
- Follow the cloud. On Google Cloud, ADK's deploy, sessions and Agent Platform integration save weeks.
- Count the tools. Many tools from many systems point to MCP support and sub-agents, which ADK, Claude Agent SDK and LangGraph handle well.
- Keep the exit open. Put business rules in plain functions and retrieval behind one tool interface, so changing frameworks later is a refactor rather than a rewrite.
We build agents on all five and have no stake in which one you pick. If you want a second opinion on an architecture, book a free strategy call and we will review it with you. For hands-on guides, see How to Build a RAG Agent with Google ADK 2.0 and Google ADK 2.0 Multi-Agent Workflows.
