AI Agent Frameworks

AI agent frameworks are software libraries that handle the plumbing of an AI agent: calling the model, exposing tools, keeping state between steps, coordinating several agents and recording what happened. Developers describe the behavior and the tools; the framework runs the loop.

In October 2026 the field splits into independent frameworks and vendor SDKs. LangGraph, from LangChain, models agents as graphs with checkpointed state and is on its 1.2 release line. CrewAI organizes agents into role-based crews and ships frequent 1.x releases. Microsoft Agent Framework, built by the AutoGen and Semantic Kernel teams as the successor to both, reached general availability in April 2026, and standalone AutoGen is now in maintenance. The model vendors ship their own: the OpenAI Agents SDK (the production successor to the experimental Swarm project), Anthropic's Claude Agent SDK (renamed from the Claude Code SDK in September 2025) and Google's Agent Development Kit, which reached version 2.0 in 2026. The Vercel AI SDK covers TypeScript and web apps, and Pydantic AI, smolagents, Agno, Mastra and AWS Strands fill other niches.

The core feature set has converged. Nearly every framework offers typed tool definitions on top of function calling, support for Model Context Protocol servers so a tool built once works across frameworks, handoffs between agents, streaming, guardrails that validate inputs and outputs, human approval steps, durable execution so long runs survive restarts, and tracing that records every model call and tool result. Differences show up in the execution model (graph, crew or handoff), language support (Google's ADK covers Python, TypeScript, Java and Go, while Microsoft's framework targets Python and .NET), and how tightly a framework assumes one model provider. Cross-agent standards such as the Agent2Agent protocol are supported unevenly.

Teams usually choose by constraint rather than by feature list: the language and cloud they already use, whether runs last seconds or days, whether they need to swap models, and how much auditability the use case demands. A thin vendor SDK is the fastest path for an app built on one provider. A graph framework with checkpointing suits regulated workflows that must pause for documents or approvals and resume exactly where they stopped. Many teams start with direct API calls and add a framework only once state, retries and tracing become hard to manage by hand.

Frameworks affect AI visibility in two ways. They are how most custom research and shopping agents get built, so their default web tools (search APIs, page fetchers, headless browsers, computer use) decide which pages those agents can actually read. A site that blocks headless browsers or renders content only client-side can be invisible to a whole class of agents. Frameworks also power internal agentic workflows for keyword research, content briefs and monitoring, often connected to analytics data through MCP. Our what are SEO agents piece covers that second use.