Manual Implementation

Before you touch any framework, you should build an agent by hand: a plain script that calls a model API in a loop, passes it tool definitions, executes whatever tool the model requests, feeds the result back, and repeats until the model produces a final answer. That’s the whole trick. An agent is a while loop around a chat completion call plus a list of functions — everything else in this module is packaging around that loop.

This matters because frameworks hide the loop, and when they break — a tool call silently dropped, context blowing past the window, an agent stuck retrying — you can only debug what you understand. Writing the agentic loop yourself demystifies tool calling, conversation state, and stop conditions, and gives you a baseline for judging every framework that follows: what is this abstraction actually saving me? For many production systems, a few hundred lines of hand-rolled orchestration beats a heavyweight dependency.

In practice you’ll install nothing but a provider SDK (openai or anthropic), define two or three tools as plain functions with JSON Schema descriptions, and write the loop: send messages, check the response for tool calls, dispatch them with a dictionary lookup, append results as tool messages, repeat. Add a system prompt, a max-iterations guard, and simple logging of every turn. Once that runs, you’ll recognize the same skeleton inside LangGraph, CrewAI, and every SDK in this module.