Part 1 of 6 · 1 min read
Overview
Modules 1 to 3 of this level were about how agents are arranged and how they exchange data. This module is about whether the arrangement ends well: an agent that never stops, stops too early, or stops for the wrong reason. Every production agent needs an answer to "who decides it is done, and what happens if it cannot be?". It is also a very common interview question.
About the example, and what is real
Everything runs offline in code/mini-goals. Real: the five stopping policies (self-report, budget, verification of claims, a stall detector with replans and escalation), the bookkeeping of outcomes, steps and cost, the Monte Carlo over 20,000 seeded tasks per policy, and the 16 tests. Assumed: the agent is a stochastic stand-in: 45 percent chance per step of completing the next subgoal, 6 percent of falling into a loop that lasts about 12 steps, 4 percent of drifting, a 5 percent chance per step of over-claiming once three subgoals are done, 15 percent of tasks impossible, a verifier that wrongly approves 2 percent of the time, a replan that frees a stuck agent 70 percent of the time, and a monitor costing 0.3 model calls per step. The numbers show how the policies transform the assumptions, not how often real agents over-claim. No model is called.