Definition
Plain language
An inference-time technique that breaks a long task into a few explicit checkpoints the agent tracks as it works.
As stated in the literature
Subgoal-Oriented planning, an inference-time scaffolding for long-horizon agents in which the same LLM generates milestone subgoals upfront and audits progress against them between actions.
Also called: Subgoal-Oriented planning
Why it matters: Explicit checkpoints reduce the chance an agent drifts off-task during long-horizon work, which is one of the main failure modes of current autonomous agents.
For example, before starting a long bug-fix task, the agent writes "1) reproduce, 2) localize, 3) patch, 4) test" and checks off each milestone as it goes.
Heard on the show
“The inference-time version they call SGO — Subgoal-Oriented planning.”Episode 008 — Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps