Most agent talks stop at the prompt. The demo works, then ships — nobody engineered what runs after turn one: the loop.
Loops aren't one thing: one that calls tools till done, one that grades and retries, one triggered by events, one that rewrites its harness from its traces. Different costs and failure modes.
We recap prompt, context, harness fast, then build a goal-condition loop live — planned failure included. You'll leave able to name, compare, and pick the right loop for the job.

Learning Outcomes:

1. Identify the loop pattern underneath any agent tool you already use
2. Compare loop types (agent, verification, event-driven, hill-climbing) on latency, token cost, and failure mode
Implement a persistent goal-condition loop with a per-turn evaluator
3. Debug a loop that fails to terminate or burns budget silently