Best AI agent frameworks in 2026, compared for production

Every framework can run the demo

Every framework here can call a model, call tools and loop until the job is done. What separates them in production is the bad day, when a tool times out mid-run, a payment needs approval, the server restarts, or someone asks why the agent did what it did last Thursday. So for each one I say what it makes easy, what it leaves to you, and when it is the wrong choice. None of them makes your writes idempotent or knows your business rules, so those are always yours.

I build and run production AI agents for operational teams, so my bias is towards frameworks that keep control flow in plain code I can read and test. No vendor has paid to be here. All nine are free to use and all but the Claude Agent SDK are open-source, so the real cost is tokens plus any hosting and tracing you buy. I checked every version, licence and price on vendor pages on 25 September 2026.

1. LangGraph

LangGraph is LangChain's orchestration library, MIT licensed, in Python and TypeScript, and stable since 1.0 in October 2025. You model the agent as a graph of steps over shared state, and a checkpointer saves that state after every step, so runs resume after a crash and can wait hours for an approval. Idempotency is yours. A step that pauses for approval reruns from the top on resume, so a write made before the pause can happen twice.

It wins when the process has real branches, loops and waits, and is the wrong choice for one agent with three tools, where the graph is ceremony. Tracing, evals and hosting are sold as LangSmith. Developer is free for one seat and 5,000 traces a month, Plus is 39 dollars per seat per month with 10,000 traces and one small deployment included, then usage is metered, and Enterprise is a custom quote.

2. CrewAI

CrewAI is an MIT-licensed Python framework with two layers. Crews are teams of role-based agents that split a task, and Flows are event-driven workflows with typed state. Its own production advice is to start with a Flow and use a Crew only where you want autonomy. Flows persist state with one decorator, pause for human feedback and send traces to CrewAI's tracing service.

It wins for Python teams that want a readable multi-agent prototype fast, and is the wrong choice for one agent following a procedure, where agents talking to each other add tokens and variance. Version 1.0 shipped in October 2025. The hosted platform publishes two tiers, Basic free with 50 workflow executions a month, and Enterprise with SSO and deployment in your own cloud on a custom quote.

3. OpenAI Agents SDK

OpenAI's Agents SDK, MIT licensed in Python and TypeScript, succeeded the experimental Swarm project. It gives you agents, handoffs, guardrails, sessions and tracing on by default. Tools can require approval, and a paused run serialises to JSON so another process can resume it. An April 2026 update added sandboxed workspaces, Python first. Durability means adding an engine, and the docs cover Temporal, Restate, DBOS and Dapr.

It wins on OpenAI models when you want traces and approvals quickly. It is the wrong choice if you need strict model neutrality, and its tracing is unavailable to organisations on zero data retention. It is still below 1.0 and its policy allows breaking changes in minor versions, so pin them. The SDK is free and you pay API prices. The Assistants API shut down on 26 August 2026, and the visual Agent Builder closes on 30 November 2026.

4. Google Agent Development Kit (ADK)

Google's Agent Development Kit is Apache 2.0 licensed in Python, TypeScript, Go and Java. Version 2.0, out for Python in May 2026, made it a graph engine with a retry policy per step, timeouts and nodes that pause for human input. The open-source kit also ships evaluation with user simulation, plus sessions, memory and tracing, and runs non-Gemini models through adapters.

It wins for Google Cloud teams who want build, eval and deploy from one vendor, and is the wrong choice if you want a small library. Moving to 2.0 meant session database changes, and the older tool confirmation feature is still experimental and does not work with persistent session services. Hosting is Agent Runtime, formerly Vertex AI Agent Engine, at 8.5 cents per vCPU-hour after 50 free hours a month and 0.9 cents per GiB-hour after 100 free, with idle time between turns not billed.

5. Microsoft Agent Framework

Microsoft Agent Framework is the MIT-licensed successor to AutoGen and Semantic Kernel, for Python and .NET, and reached 1.0 on 3 April 2026 with a long-term support commitment. It gives you graph workflows with checkpointing, human approval and pause and resume across its orchestration patterns, OpenTelemetry tracing, and connectors for Azure OpenAI, OpenAI, Anthropic, Bedrock, Gemini and Ollama.

It wins for .NET shops and Azure estates, where identity, monitoring and hosting match what IT already runs. It is the wrong choice for TypeScript teams, or small teams avoiding Azure, since off Azure the checkpoint storage, hosting and trace collection are yours to run. With AutoGen frozen and Semantic Kernel on a support clock, new work belongs here. Hosted agents in Foundry Agent Service cost 9.94 cents per vCPU-hour and 1.18 cents per GiB-hour in East US.

6. Claude Agent SDK

Anthropic's Claude Agent SDK, renamed from the Claude Code SDK in September 2025, packages the agent loop and tools behind Claude Code, in Python and TypeScript. You get file editing, shell, web search, subagents, MCP, hooks at each step, and permissions that decide which tools need a person. For agents working inside files, documents and code, it is the shortest path here to a capable agent.

It leaves operations to you. Each session is a subprocess with its transcript on local disk, so without a session store you lose history on restart, and you host and sandbox the container yourself. It is the wrong choice if you need other models or an open-source licence, because it runs only Claude under Anthropic's commercial terms. Managed Agents, in public beta since April 2026, hosts the loop for you at token prices plus 8 cents per session-hour of running time.

7. Pydantic AI

Pydantic AI is the MIT-licensed Python framework from the team behind the Pydantic validation library. It is typed end to end, so your type checker catches mistakes in tools, dependencies and outputs before production does, and swapping models is a one-line change. Rather than being a durable engine, it plugs into seven, including Temporal, DBOS and Prefect, with tool approval built in and Pydantic Evals for testing agent behaviour.

It wins for Python teams that think in types and already trust a durable engine. It is the wrong choice for TypeScript teams or anyone wanting an all-in-one platform, and it moves fast, with V1 in September 2025 and V2 in June 2026. The framework is free. Logfire, its observability product, has a free personal plan with 10 million records a month, a Team plan at 49 dollars a month and Growth at 249 dollars a month.

8. Vercel AI SDK

Vercel's AI SDK is an Apache 2.0 TypeScript toolkit for model calls and streaming, and version 7, released on 25 June 2026, made it an agent runtime. It added tool approvals, timeouts per step and per tool, OpenTelemetry, and a WorkflowAgent that survives restarts and deploys by running on the open-source Workflow SDK, which works locally, on Postgres or on Vercel.

It wins for product teams building agents into a web app. It is the wrong choice for Python teams, or when you want a prescribed multi-agent design, because it gives you primitives and leaves architecture, idempotency and evals to you. The SDK is free and runs anywhere. Vercel's Hobby plan includes 50,000 workflow events a month, with on-demand usage at 2 cents per 1,000, and its AI Gateway adds no markup to provider prices.

9. Mastra

Mastra is a TypeScript framework, Apache 2.0 licensed apart from separately licensed enterprise features, and stable since 1.0 in January 2026. It bundles agents, typed workflows that suspend and resume from snapshots in your own database, memory, scorers that grade outputs live or over stored traces, and a studio for inspecting runs.

It wins for TypeScript teams who want LangGraph-style workflows, memory and evals without leaving Node. It is the wrong choice for Python shops, and for a plain tool loop the Vercel AI SDK alone is lighter. The framework is free, and you run its storage and hosting unless you use the hosted platform, which is free for 100,000 observability events and 24 CPU hours a month, with Teams at 250 dollars a month and Enterprise on a custom quote.

What changed in 2025 and 2026

AutoGen and Semantic Kernel are folded into Microsoft Agent Framework, the Claude Code SDK is now the Claude Agent SDK, LangGraph Platform is LangSmith Deployment, and Vertex AI Agent Engine is Agent Runtime on Gemini Enterprise Agent Platform. OpenAI's Assistants API is gone, and LlamaIndex deprecated its TypeScript library in March 2026.

Most of the field also reached a stable or second major version in the last year, so tutorials older than that often show APIs that no longer exist. The OpenAI and Anthropic agent SDKs are still below 1.0.

Three names I left off the list

LlamaIndex still maintains Python Workflows but now calls itself a document processing platform, so use it for parsing messy documents, with paid LlamaParse plans from 50 dollars a month, rather than as your agent runtime. DSPy, MIT licensed and on version 3, tunes prompts against a metric and examples you supply, which pays off once you have a labelled set. smolagents from Hugging Face is a small Apache 2.0 library whose agents write Python as their actions, good for experiments, but code-writing agents need a sandbox and a reviewer.

What matters more than the framework

In my experience the framework is rarely why a production agent fails. Four things decide it, and no library does them for you.

The first is a deterministic rules layer next to the model, so plain code checks that totals add up, suppliers exist and discounts stay within limits before anything is written. The model proposes and the rules decide. The second is evals built from real past runs, where last quarter's actual inputs, awkward ones included, are replayed against the answers a person accepted on every prompt or model change. The dataset is the asset, and only you have it.

The third is a human owner for the decision, a named person who approves the risky actions and answers for the outcome, with approval gates on what is expensive to reverse rather than on everything. The fourth is an undo path designed before the first write, with drafts before sends, reversible postings, an audit log and idempotency keys, because durable execution replays steps, and a replay without idempotency pays the same invoice twice.

Comparison at a glance

How to pick

Start from your language and your cloud. In Python, choose LangGraph for processes with real branches and waits, or Pydantic AI for typed code on a durable engine you already run. In TypeScript, choose the Vercel AI SDK for agents inside a product, or Mastra for workflows, memory and evals in one package. On Azure or .NET choose Microsoft Agent Framework, and on Google Cloud, ADK. If you are committed to one model vendor, its own SDK is the shortest path.

A small team with one workflow should not adopt a framework for its own sake, and sometimes a script beats an agent. Adopt one when you need durable approval pauses or a trace an auditor can read. Larger organisations should standardise on one framework per language and spend the saved effort on evals and rules.