LangGraph
A low-level orchestration framework for building resilient, stateful agents with durable execution, human-in-the-loop control, and long-term memory. Built by LangChain, usable standalone.
Explore14 Frameworks · Checked October 2026
CrewAI
A framework for orchestrating role-playing, autonomous AI agents, empowering them to collaborate as a crew to tackle complex, multi-step tasks with both no-code and code-first options.
ExploreOpenAI Agents SDK
A lightweight yet powerful, provider-agnostic framework for building multi-agent workflows, with built-in tracing, guardrails, sessions, and support for 100+ LLMs.
ExploreGoogle Agent Development Kit (ADK)
An open-source, code-first Python toolkit for building, evaluating, and deploying sophisticated AI agents. Optimized for Gemini but model-agnostic and deployment-agnostic.
ExplorePydantic AI
A GenAI agent framework built by the Pydantic team, bringing FastAPI-style ergonomics to agent development with type safety, structured outputs, and seamless observability.
Exploresmolagents
A barebones Hugging Face library for agents that think in code — the core logic fits in ~1,000 lines, with first-class support for code-writing agents and sandboxed execution.
ExploreMastra
A TypeScript framework for building AI-powered applications and agents, with model routing, graph-based workflows, human-in-the-loop, memory, and built-in MCP server support.
Explore Note: TypeScript-based, not Python — included for its prominence in the agentic framework landscape.Microsoft Agent Framework
An open, multi-language framework for building production-grade AI agents and multi-agent workflows in Python and .NET, with graph-based orchestration and enterprise observability.
ExploreStrands Agents
A model-driven SDK for building AI agents in just a few lines of code, with Python and TypeScript SDKs, native MCP support, and model-agnostic providers from Bedrock to Ollama.
ExploreLlamaAgents
An event-driven, async-first, step-based framework from the LlamaIndex team for controlling execution flow in document-centric agent applications, from library to deployable service.
ExploreAutoGen
Microsoft's multi-agent programming framework: AutoGen Core is an event-driven actor model and AgentChat adds round-robin, selector, Magentic-One and Swarm team patterns. It is now in maintenance mode, community-managed, with no new features; Microsoft points new projects to Microsoft Agent Framework.
Note: maintenance mode — start new projects on Microsoft Agent Framework (see the migration guide on Microsoft Learn). ExploreLlamaIndex
A data-centric framework combining tools, memory, RAG components, multi-agent handoffs and event-driven Workflows. A strong fit for enterprise search, document synthesis and analytical agents.
ExploreSemantic Kernel
Microsoft's enterprise SDK for C#, Python and Java with plugins, agents and an event-driven Process Framework. Microsoft describes Microsoft Agent Framework as its successor, so check the migration guide before starting new work.
Note: succeeded by Microsoft Agent Framework (card 08). ExploreOGX (formerly Llama Stack)
The renamed Llama Stack project: a model-agnostic, OpenAI-compatible agentic API server with tool calling, MCP integration and file search. The GitHub repository now lives at ogx-ai/ogx.
ExploreChoosing a Framework · summarised from ODSC (Sep 2026)
Pick the architecture first
There is no universally best framework. Decide where state must persist, where model reasoning may stay non-deterministic, where humans must approve actions and how failures are recovered, then choose a library.
Match the framework to the job
LangGraph when checkpointed, explicit control dominates; AutoGen-style conversational coordination when dynamic multi-agent delegation dominates (for new work, via Microsoft Agent Framework); CrewAI when development speed and role clarity dominate; LlamaIndex when retrieval and data reasoning dominate; Microsoft Agent Framework or Semantic Kernel in Microsoft-centred stacks.
Five criteria for production
State and memory, orchestration, tool integration, governance and observability. The recurring trade-off is abstraction versus control: role-based frameworks are quick but less transparent than an explicit graph.
Plan for agent security
The OWASP Top 10 for Agentic Applications 2026 covers goal hijacking, tool misuse, identity and privilege abuse, supply-chain weaknesses, unexpected code execution and memory or context poisoning. Use least-privilege tools, approval gates, strict schemas, sandboxed execution and tracing of tool calls.
OWASP Top 10 for Agentic Applications 2026Evaluate whole trajectories
Test complete runs, not only final answers: right tools chosen, constraints respected, errors recovered and a correct stop. Cost and latency compound across loops, so set termination conditions and message limits.
ReAct paper (reasoning and acting)Standards that connect frameworks
The July 2026 Model Context Protocol specification added a stateless core, authorization updates, formal extensions and JSON Schema for tools, which helps interoperability across these frameworks.
MCP specification