One page to shortlist by. Pair with the decision guide in Chapter 3 and confirm with a two-day prototype on your own workload.
Pick this When Mind the LangGraph complex branching, audits, pause/resume, durable state learning curve; graph thinking required LangChain fast prototyping, big integration surface abstraction churn; graduate to LangGraph CrewAI process maps cleanly to roles (research → write → review) role overhead on simple linear tasks OpenAI Agents SDK all-in on OpenAI; want speed and clean tracing vendor coupling Claude Agent SDK desk-work agents: coding, files, terminal, sub-agents Anthropic-centric by design Google ADK multi-agent hierarchies; native A2A; Google Cloud estate assumes Google tooling Pydantic AI type-safety, testability, production validation first smaller ecosystem than the giants smolagents minimal code-acting agents; research; HF ecosystem code execution needs sandboxing MS Agent Framework Microsoft / .NET estates; SK + AutoGen successor newest of the set; migration from SK/AutoGen ongoing AG2 conversational multi-agent research and dialogues community fork; check maintenance fit LlamaIndex document-centric agents and agentic RAG less suited to general orchestration Haystack production search + RAG pipelines with agent steps pipeline mindset, not free-form autonomy
Low-code tier (n8n, Dify, Copilot Studio) fits linear, low-risk internal automations. TypeScript-first teams: Mastra. And for many production systems the honest answer remains ~100 lines of your own loop (Chapter 2) plus MCP tools — add a framework when you feel the ceiling, not before.