AI agent frameworks compared: which one to build on
A coding agent writes your code. An agent framework is what you reach for when the agent is your product: it runs the loop of model calls and tool calls, keeps state, and lets a human step in. Most of these now work with more than one model provider, so the real differences are language, how much control you get over each step, and which ecosystem you already use. Licenses below come from each project's GitHub repo. We left out stars and pricing because they change every week.
OpenAI's lightweight framework for multi-agent workflows. Agents get instructions, tools, guardrails and handoffs, with built-in sessions, human-in-the-loop and tracing. It also covers realtime voice agents and sandbox agents that work inside a container. Its README says it is provider-agnostic: it supports OpenAI's Responses and Chat Completions APIs plus 100+ other LLMs.
Our take, best for: A small API with tracing built in, especially for voice agents.
Anthropic describes it as Claude Code as a library. Your agent gets Claude Code's tools for reading and editing files and running commands, plus MCP, subagents, skills, hooks and permission controls. The Python package bundles the Claude Code CLI, so there is nothing else to install. It used to be called the Claude Code SDK. The Python repo is MIT-licensed; the docs link Anthropic's license and terms.
Our take, best for: Agents that work on files and code: internal dev tools, research bots, ops agents.
Google's open-source, code-first toolkit, now at version 2.0 in Python. It adds a graph-based workflow runtime (routing, fan-out, loops, retries, human-in-the-loop), structured agent-to-agent delegation, and MCP and OpenAPI tools. It is optimized for Gemini but model-agnostic, and deploys to Cloud Run or Vertex AI Agent Engine.
Our take, best for: Teams on Gemini or Google Cloud, or anyone who wants one framework across several languages.
An open-source SDK for building and running agents inside your own process, with no hosted control plane. It supports Amazon Bedrock, Anthropic, OpenAI, Gemini and other providers. The repo (now strands-agents/harness-sdk) also ships a ready-made harness agent, a CLI, and plugins for Claude Code, Codex and Cursor.
Our take, best for: Builders who would otherwise hand-roll their own agent loop, especially on Bedrock.
LangChain's low-level orchestration framework for long-running, stateful agents. It offers durable execution that resumes where it left off after a failure, human-in-the-loop interrupts, and short- and long-term memory. For a quicker start, LangChain points to Deep Agents, a higher-level package built on LangGraph.
Our take, best for: Complex flows where you want to control every step and survive crashes.
An open-source Python framework with two layers: Crews, teams of role-based agents that collaborate, and Flows, event-driven workflows for precise control. A paid AMP Suite adds managed deployment, observability and governance.
Our take, best for: Getting a multi-agent team (researcher, writer, reviewer) working fast.
Microsoft's framework for production agents and multi-agent workflows, with graph-based patterns (sequential, concurrent, handoff, group collaboration), middleware, durability and observability. Version 1.0 is a production release with stable APIs, and Microsoft calls it the enterprise-ready successor to AutoGen.
Our take, best for: .NET shops, and teams on Azure or Microsoft Foundry.
A free, provider-agnostic TypeScript toolkit for AI apps and agents. It works with Next.js, React, Svelte, Vue, Angular and Node.js. By default it routes model names through Vercel AI Gateway, or you can install a provider package to connect directly.
Our take, best for: Adding an agent or chat feature to a web app you already ship.
A TypeScript framework for agents and AI apps: agents, a graph-based workflow engine, suspend-and-resume human-in-the-loop, memory, RAG, evals and observability, and routing to 40+ model providers through one interface. It can also publish your agents and tools as MCP servers.
Our take, best for: TypeScript teams who want workflows, memory and evals in one package.
From the Pydantic team: a typed agent loop where switching models is a one-string change. Companion packages add a harness (memory, guardrails, sub-agents), evals that grade tool-call trajectories, and graphs.
Our take, best for: Python devs who want type-checked inputs and outputs and testable agents.
Hugging Face's minimal agent library: the core logic fits in about 1,000 lines. Its CodeAgent writes its actions as code, which you can run in sandboxes such as E2B, Modal, Blaxel or Docker. It works with any model, including local ones, and can use tools from MCP servers or the Hugging Face Hub.
Our take, best for: Learning how agents work, and experiments with open or local models.
Trace from day one. Turn on tracing so you can see every model call and tool call in a run. The OpenAI Agents SDK has it built in. Source
Put a human on risky tools. Require confirmation before tools that spend money, send messages or delete things. ADK has a tool confirmation flow for exactly this. Source
Evaluate the path, not just the answer. Check which tools the agent called and in what order, not only its final output. Pydantic Evals grades tool-call trajectories. Source
Connect tools with MCP. If your tools are MCP servers, you can switch frameworks without rewriting them. Source
FAQ
What is an AI agent framework?
A library for building software where a model decides which tools to call, in a loop, until a task is done. Frameworks handle the loop, tool calling, memory and state, and usually tracing and human approval steps.
Do I need an agent framework at all?
Our take: not always. One model with a few tools works fine with a lab SDK or even a hand-written loop. Reach for an orchestration framework when runs are long, have many steps, need approvals, or must resume after a crash.
What happened to Microsoft AutoGen?
Microsoft has put AutoGen in maintenance mode, limited to bug fixes, security patches and docs. It recommends new users start with Microsoft Agent Framework and publishes a migration guide.
Which agent frameworks support MCP?
The OpenAI Agents SDK, Claude Agent SDK, Google ADK and smolagents all list MCP tools in their docs, and Mastra can publish agents and tools as MCP servers. Check each project's docs for the transports it supports.
What's the best agent framework for TypeScript?
Our take: the AI SDK if you are adding AI features to a web app, Mastra if you want workflows, memory and evals in one package. The OpenAI Agents SDK, Claude Agent SDK, ADK, LangGraph and Strands also have TypeScript versions.
How we write these: each tool description is based on the tool's own site or docs (linked as Source) and checked on the date above. "Our take" lines are our opinion. Something out of date? Email hello@hotfixdaily.com. No tool paid to be listed.
Get the next issue in your inbox
Five things worth your attention in AI, agents, vibe coding and game dev, every weekday. Plus the free AI Builder Starter Stack PDF when you join.
