Definition
Plain language
A deployment trick where an AI agent gets free time to wander before being given the actual task.
As stated in the literature
A two-phase agent deployment pattern in which the agent first performs unconstrained environment exploration, summarizes findings into a natural-language note, and then receives the task with that note injected into the prompt.
Why it matters: Letting agents build up environment knowledge before being graded dramatically improves their performance on unfamiliar tools without retraining.
For example, before asking an agent to file a bug report, you let it click around the issue tracker for a minute, write notes about the interface, and then give it the actual task with those notes attached.
Heard on the show
“So they propose what they call Explore-then-Act, which is a deployment pattern.”Episode 052 — An Old Reinforcement Learning Tradeoff Sneaks Back Into LLM Agents