Agentic Loop Engineering
An eighteen-notebook course that measures each component of the recursive loop behind an AI coding agent, such as verification and state persistence, with reproducible numbers on real benchmarks.
This course isolates each component of what it calls loop engineering, the design of the recursive system that discovers work, assigns it to an agent, verifies the result, persists state and decides the next action until a goal is met or a human is needed. Each of its eighteen notebooks takes one component, shows a naive baseline failing with real measured numbers on industry-standard software datasets, then shows that component closing the gap and hardens it against its documented failure mode.
Its stated central claim is that a loop is only as good as the verifiable signal it is wired to: a loop that only re-runs an agent against its own opinion barely improves, while one wired to an executable check, such as a test, a schema or a real evaluation harness, measurably does. All results were produced on a single A100 GPU with a self-hosted open model, captured inline in the notebooks and traceable to a saved JSON results file, so the whole flow can be followed by reading the saved output without rerunning anything.