AI agent frameworks compared

Write the loop yourself for a first agent, and adopt a framework when you need something specific that the loop does not give you: saved state, approval pauses, several agents working together, or built-in tracing. A framework does not make an agent smarter. It saves you from writing infrastructure around the same model calls.

The short version

• The raw SDK is enough for a single agent with a handful of tools. See the Python walkthrough.
• Frameworks add infrastructure: persistence, approvals, handoffs, tracing, ready-made tools.
• The cost is a layer of abstraction between you and the requests actually sent to the model.
• If the job is coding, skip both. A finished coding agent already exists.

Raw SDK or framework

Anthropic’s guide to building effective agents is blunt about the order: “We suggest that developers start by using LLM APIs directly: many patterns can be implemented in a few lines of code.” It adds that if you do use a framework, you should understand the code underneath it.

That advice holds because the two options differ in what you maintain, not in what the model can do.

You need	Raw SDK loop	A framework
One agent, a few tools	A short loop you fully understand	More machinery than the job needs
To pause and wait for a person to approve a step	You design and store the paused state yourself	Usually built in
To survive a crash halfway through a long run	You write the checkpointing	Usually built in
Several agents handing work to each other	You write the routing	A core feature of most frameworks
A record of every model call and tool call	You add logging	Usually built in as tracing
To see the exact request sent to the model	Always visible, because you wrote it	Sometimes hidden behind the framework’s own prompts

A reasonable rule: start a framework evaluation only when you can name the row you need.

Seven agent frameworks and SDKs

Each row describes the project in terms taken from its own repository or documentation, linked in the first column. Languages and licenses were read from those sources on 13 November 2026. Where a project keeps one repository per language, the license shown is the one on the linked repository. Check again before you depend on one.

Framework	From	Languages	License	What it is for
Claude Agent SDK	Anthropic	Python, TypeScript	Python repository: MIT. TypeScript repository: Anthropic’s Commercial Terms of Service.	Claude Code as a library: the same agent loop with built-in tools for files, commands and the web, plus permissions, sessions, subagents and MCP.
OpenAI Agents SDK	OpenAI	Python, TypeScript	MIT	A lightweight framework for multi-agent workflows: agents, handoffs between them, guardrails, sessions and tracing. Its README describes it as provider-agnostic.
LangGraph	LangChain	Python, TypeScript	MIT	A low-level orchestration framework for long-running, stateful agents, with durable execution and human-in-the-loop pauses.
CrewAI	CrewAI	Python	MIT	Teams of role-playing agents, called Crews, combined with event-driven Flows for the parts you want to control precisely.
Pydantic AI	Pydantic	Python	MIT	A typed agent loop from the team behind Pydantic, built to work across model providers.
Agent Development Kit (ADK)	Google	Python, Java, Kotlin, Go, TypeScript	Apache-2.0	A code-first toolkit for building, evaluating and deploying agents. Its README says it is optimized for Gemini and model-agnostic.
Microsoft Agent Framework	Microsoft	Python, .NET	MIT	Building, orchestrating and deploying agents and multi-agent workflows.

This table is a description, not a ranking. We have not benchmarked these projects against each other, and the right choice depends more on your constraints than on any feature count.

How to choose

Work down this list and stop at the first line that applies.

1. You are building your first agent. Use none of them yet. Write the loop.
2. Your agent’s job is to work on files and run commands, the way a coding agent does. The Claude Agent SDK gives you those tools already built, which is most of the work.
3. You are committed to one model provider. Use that provider’s SDK: the Claude Agent SDK, the OpenAI Agents SDK or Google’s ADK. The defaults will match the API you are calling.
4. You need long-running, resumable runs with explicit control over each state. That is the problem LangGraph describes itself as solving.
5. You want typed inputs and outputs in a Python codebase. Look at Pydantic AI.
6. You are modelling a team of specialists. CrewAI and the OpenAI Agents SDK are both organised around several agents. Confirm first that one agent with more tools cannot do the job, because every added agent adds cost and failure modes.
7. Your stack is .NET or Java. Microsoft Agent Framework and ADK cover those languages.

Whatever you pick, keep your tools as plain functions with clear descriptions. They are the part of your agent that survives a change of framework. Tools and MCP explains how to write them.

What a framework costs you

• Hidden prompts. Anthropic’s guide notes that frameworks often add layers of abstraction that can obscure the underlying prompts and responses. When an agent misbehaves, you need to see the exact text the model received, so learn how to print it.
• A moving target. These projects release often. Pin the version you tested against and read the changelog before upgrading.
• Temptation. Anthropic’s guide warns that frameworks make it easy to add complexity when a simpler setup would do. Multi-agent features are the usual example.
• A second thing to debug. A failure can now be in your code, the framework or the model.

None of this is a reason to avoid frameworks. It is a reason to adopt one for a named need and to keep a test set that tells you whether the move helped.

Hosted agent runtimes

A third option sits beside frameworks: let the model provider run the loop on its own servers. Anthropic’s documentation describes Managed Agents as a hosted agent harness that runs the agent loop, and OpenAI’s agents guide describes its Agents API as the choice for long-running tasks where OpenAI manages the agent. You write less code and accept more dependence on one vendor. Read the provider’s documentation for current availability and terms.

Our pick · we make it

If the job is coding, you may not need a framework at all

Emdash is our product, and it is not an agent framework. It does not help you write an agent. It is for the case where the job you had in mind is software development, which coding agents such as Claude Code and Codex already do. Emdash runs those agents in parallel, each task in an isolated workspace, with diffs and pull requests to review.

Emdash Classic is a free, open-source (Apache-2.0) desktop app for macOS, Windows and Linux, and you bring your own agent subscription or API key. Emdash 2.0, which runs agents in the cloud for teams, is in early access. If you are building an agent for anything other than coding, use one of the frameworks above or the raw SDK.

See Emdash →

Frequently asked questions

Do I need a framework to build an AI agent?

No. An agent is a loop around a model API, and you can write that loop in well under a hundred lines with a provider's official SDK. A framework earns its place when you need features the loop lacks, such as saving state between steps, pausing for human approval, or passing work between several agents.

Which AI agent framework is best for beginners?

For learning, none. Write the loop by hand first so you understand what a framework does for you. After that, pick the SDK published by the model provider you already use, because its defaults match that provider's API and its documentation is maintained alongside it.

Are AI agent frameworks free?

Most are open source under the MIT or Apache-2.0 license, so the library costs nothing. You still pay the model provider for every request the framework makes on your behalf, and any hosting or monitoring service you add has its own pricing.

Can I switch frameworks later?

Yes, but the cost depends on how much of your agent lives in framework-specific code. Tool functions and prompts usually move with little change. Orchestration, saved state and tracing are tied to the framework and have to be rebuilt, which is one reason to keep your tools as plain functions.