Appendices

Chapter A

Framework Quick Reference

A one-page AI agent framework shortlist covering production fit, strengths, and tradeoffs.

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.