Definition
Plain language
A helper network that estimates how far along an AI agent is toward finishing a task, used to give it progress-based encouragement.
As stated in the literature
In MiRA, a learned regression model predicting the fraction of subgoals completed at a given state; its temporal differences supply a potential-based shaping reward, kept separate from the value critic that learns only from the final task outcome.
Why it matters: By rewarding visible progress instead of only the final result, it helps an agent learn faster on long tasks where success comes only at the very end.
For example, partway through a multi-step task it estimates the agent is about 60% done and nudges it with encouragement for making progress.
Heard on the show
“They call it the potential critic.”Episode 008 — Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps