Closing: Frameworks Are Not the Skill

You can build a working agent in 50 lines of Python. You can build a production agent in 5,000. You leave this course knowing the difference. The pattern: framework choice matters less than you think, design choices matter more than you think. LangChain, LangGraph, CrewAI, AutoGen, building from scratch. Any of them work. None of them save you from the hard parts: tool definitions that the model can use reliably, state management that survives errors, evaluation that catches regressions, observability that lets you debug at 2 a.m. The weXare thesis: the best agent is the one humans can supervise meaningfully. That means clear tool boundaries, deterministic behavior where possible, traceable decisions, kill switches that work. The agent is a colleague, not a black box. **Five takeaways to keep:** 1. Start with LangChain for simple cases. Move to LangGraph when state, durability, or multi-agent matters. 2. Tool definitions are the bottleneck in agent quality. Spend time here, not on prompt tuning. 3. State management is the second bottleneck. Durable state survives reality. 4. Evals are not optional. Without them, you are shipping vibes. 5. Anthropic principles: simple agents, explicit tools, durable workflows, human checkpoints. **What is next:** Take [LLMs in Production](/en/learn/llms-in-production) for the operational side. Take [Advanced HITL Patterns](/en/learn/advanced-hitl-patterns) for the design discipline of human oversight. Take [Scaling Human-Centered AI](/en/learn/scaling-human-centered-ai) when you scale beyond one team. Now go ship an agent that does not embarrass you in production.