Definition
Plain language
An AI agent design that breaks tasks into explicit milestones and uses them to guide and reward progress.
As stated in the literature
A long-horizon agent architecture using LLM-generated subgoals both as inference-time checkpoints and as a basis for dense reward shaping during RL training.
Why it matters: Dense milestone rewards make long-horizon tasks trainable where a single success-or-failure signal at the end would be too sparse to learn from.
For example, an agent told to 'book a trip' might generate subgoals like 'find flights,' 'compare hotels,' and 'reserve rental car,' and get partial credit as it hits each one.