The landscape

AI agent frameworks: the landscape and how to choose

What an agent framework gives you

An agent is, at its core, a model calling tools in a loop on feedback from its environment. You can write that loop by hand. A framework exists to give you the recurring scaffolding so you don't:

• The agent loop — send a request, read the stop reason, run any requested tool, feed the result back, repeat until done, bounded by a budget.
• Tool wiring — turning functions or APIs into schemas the model can call reliably, often via the Model Context Protocol.
• Memory and state — working context now, plus durable state that outlives a single run.
• Multi-agent coordination — handoffs, sub-agents, and routing between specialised agents. See multi-agent orchestration.
• Operational hooks — persistence, human-in-the-loop checkpoints, tracing, and guardrails around the loop.

The tradeoff is the same one every abstraction makes: convenience now in exchange for a layer between you and the model. The frameworks below sit at different points on that control-versus-convenience line, which is the first axis to judge them on.

The main agent frameworks

These are the libraries a senior team is most likely to be choosing between in 2026. Positioning is neutral and use-case based — not a popularity ranking, because star counts and "fastest-growing" labels shift month to month and disagree across sources.

A few notes the table flattens. LangGraph reached v1.0 in late 2025 with a stability commitment, and LangChain's higher-level agent abstractions now build on top of it. AutoGen's lineage split: Microsoft's supported successor merges AutoGen and Semantic Kernel into the Microsoft Agent Framework, while AG2 is the independent fork from AutoGen's original authors. The Claude Agent SDK was renamed from the Claude Code SDK in 2025 as its scope broadened from coding to general agents. Strands and ADK both default to their parent cloud's models but are deliberately provider-agnostic, so neither forces a single LLM.

Framework vs no framework

The most underrated option is the last row of the table. Anthropic's own guidance on building effective agents is blunt about it: many agent patterns are a few lines against the raw model API, and a framework's abstraction can hide the exact prompts and tool definitions you most need to see when something misbehaves.

• Reach for no framework when the system is a single agent with a handful of tools, when you need to see and tune every prompt and tool schema, or when an extra dependency's churn is a liability. Starting direct also teaches you what the frameworks are actually doing.
• Reach for a framework when you need genuine multi-agent coordination, durable/long-running execution with persistence and human-in-the-loop, or when its ecosystem (tracing, integrations, deployment) saves real time. Even then, the advice is to keep abstraction thin and add it only where it pays.

This site is an instance of the no-framework choice: its coach is a hand-written bounded loop, not a framework agent. The reasoning is in why the coach runs a bounded agentic loop.

How to choose

Frameworks converge on similar capabilities, so the decision is mostly about fit. Four questions resolve most cases:

• Control vs convenience. Want fine-grained control over the graph, state, and every prompt? LangGraph or no framework. Want to stand a crew up fast with sensible defaults? CrewAI or the vendor SDKs.
• Is it really multi-agent? Most "multi-agent" needs are one agent with good tools. If you genuinely need specialised agents with isolated context, weigh CrewAI, AG2, ADK, or LangGraph's subgraph model — and read multi-agent orchestration first.
• Which ecosystem are you in? Deep on AWS, Strands; on Google Cloud, ADK; committed to Claude, the Claude Agent SDK; on Azure/Microsoft, the Microsoft Agent Framework. Matching your platform cuts integration work — see the Claude Agent SDK guide for that path.
• How much lock-in can you accept? Provider-agnostic SDKs (Strands, ADK, LangGraph) keep the model swappable; tighter vendor SDKs trade that for a smoother first-party experience. Decide deliberately, because migrating an agent's orchestration later is expensive.

The frameworks are not the architecture — they implement it. If you can't yet name the layers a framework is assembling for you, start with agentic AI architecture and choose a framework second.

Frequently asked questions

What is an AI agent framework?

It's a library that provides the reusable scaffolding for building an agent: the model-call loop, tool wiring, memory and state, multi-agent coordination, and operational hooks like persistence and tracing. It saves you rebuilding that plumbing for every project, at the cost of a layer of abstraction between your code and the model.

What is the best AI agent framework?

There isn't one. LangGraph, CrewAI, AutoGen/AG2, and the Claude, OpenAI, Google ADK, and AWS Strands SDKs all ship production agents; the right pick depends on how much control you need, whether the work is truly multi-agent, which cloud or model ecosystem you're in, and your tolerance for lock-in. For many single-agent systems, no framework is the best answer.

What are the main agent frameworks?

As of mid-2026 the commonly evaluated options are LangGraph (LangChain), CrewAI, AutoGen and its community fork AG2, the Claude Agent SDK (Anthropic), the OpenAI Agents SDK, Google's Agent Development Kit (ADK), and AWS Strands. Building directly on a model API with your own tool-call loop — no framework — is a legitimate option alongside them.

Do you need a framework to build an agent?

No. An agent is fundamentally a model calling tools in a loop, which you can write directly against the model API in a small amount of code. Frameworks help when you need multi-agent coordination, durable long-running execution, or a mature ecosystem. For a single agent with a few tools, going framework-free keeps prompts and tool schemas fully visible and avoids extra dependency churn.

LangGraph vs CrewAI vs AutoGen?

LangGraph is low-level and graph-based, favouring fine-grained control over state and durable execution. CrewAI is higher-level and role-oriented, optimised for standing up collaborating "crews" quickly. AutoGen pioneered conversational multi-agent patterns but is now in maintenance; its active community lineage is AG2, while Microsoft's supported successor folds AutoGen into the Microsoft Agent Framework. Choose LangGraph for control, CrewAI for speed of assembly, and AG2 for conversational multi-agent research.

Which agent framework should a beginner use?

Start by building one agent directly against a model API so you understand the loop, then adopt a framework when you hit a real need. When you do, a higher-level framework like CrewAI, or the SDK matching your model (the Claude or OpenAI Agents SDK), has the gentlest on-ramp; LangGraph offers more power but a steeper curve. Learn the architecture first so the framework is a convenience, not a black box.