Agents the Hard Way

Why it matters

Every framework option for a reasoning loop — step limits, parsing recovery, tool-error handling — wraps a few lines of plain code. Writing those lines once makes it clear what a framework controls for you and what it cannot.

What to practise

The notebook has you assemble four parts with the OpenAI client against NVIDIA’s hosted endpoint: the model, an add tool described as a JSON schema, a message list as short-term memory, and routing that you perform by hand. Decode each entry in tool_calls, run the tool, refuse an unknown tool name, and append the result as a tool message with its tool_call_id. Watch that the API remembers nothing between calls: the message list is the whole state.

How it fits

This is the single iteration that 1.2.2 bounds: a complete runner repeats it until it reaches an answer or a step limit. The native tool calling you route by hand here is one of the three styles that 1.2.3 compares with ReAct and ReWOO.