When faced with a hard decision, people often perform simple actions to understand the situation better—turning an object to see its other side, placing alternatives side by side, or altering a condition and observing the outcome. These actions do not complete the task themselves, but they improve the evidence needed for the next choice. LLM‑based agents can search and explore, yet current designs pay little attention to an earlier question: is the available evidence ready for a decision?
Missing evidence, evidence whose format hides what matters, or the lack of a necessary comparison can all block a sound choice. Cognitive science calls actions that improve the basis for a later decision epistemic actions. We bring this notion to LLM‑based agents and distinguish three modes:
- Acquiring missing evidence – actively seeking or requesting external information to fill gaps.
- Transforming available evidence – reformatting, visualising, or otherwise processing existing data so it becomes easier to compare and evaluate.
- Probing a system to create a revealing response – asking targeted questions or interacting in ways that elicit hidden reasoning or latent answers.
We term the interfaces, tools, and environments that enable these actions epistemic scaffolding, emphasizing that they should be auditable. The paper argues that agent design must explicitly address how decision‑ready evidence is produced, not merely focus on the final decision algorithm.
Review: Introducing epistemic actions into LLM agents offers a systematic way to boost decision reliability, and it merits further experimentation in real‑world systems.