Building agents from scratch is simple for the first demo and tedious after that. The hard parts are everything around the model call: tool dispatch, state management, retries, tracing, human approval, memory, evaluation hooks, and making the loop stop for the right reason.

Agent frameworks provide reusable abstractions for common infrastructure: tool calling, graph execution, checkpointing, multi-agent handoffs, observability, and workflow orchestration. They can save time, but they also impose an execution model. A framework is a trade-off. It does not remove the need to understand how agents work.

This chapter compares common framework styles and gives you a decision framework for when to use one, when to build it yourself, and what patterns you should understand regardless of the library.