Tool calling, planning, multi-step execution, memory, and the failure modes that only appear in loops.
10h · 8 lessons · 1 challenge
Before this: Retrieval-Augmented Generation
An agent is a model in a loop with tools, and everything hard about it comes from the loop: compounding errors, unbounded cost, silent failure, and recovery. This module builds one properly, with the guardrails and observability that make it operable.
The approval-gate pattern from our own application assistant is one of the worked examples — irreversible actions should require a human, and this is how that is built.
An agent is burning budget and going nowhere. Read the trace, find the step where it lost the plot, and place one guardrail.
Worth 600 XP.
