A new paper argues that AI agents often act on evidence they haven't actually checked.
Researchers behind "Before Agents Decide: Epistemic Action in LLM-Based Systems," posted to arXiv on October 2, 2026, say agent builders have focused on giving LLM-based systems the ability to search and explore, but have paid less attention to a more basic question: is the evidence an agent already has actually good enough to decide on? Sometimes the evidence is simply missing. Other times it exists but is in a form that obscures what matters, or there's no comparison yet that would let the agent judge it properly. The paper borrows the term "epistemic action" from cognitive science - the actions people take not to finish a task, but to sharpen their own judgment, like turning an object over to see its other side. It sorts these into three categories for agents: acquiring evidence that's missing, transforming evidence into a more useful shape, and probing a system to force a more revealing response. The authors call the tools and environments that support this "epistemic scaffolding."
Most agent research optimizes for task completion - did the agent book the flight, close the ticket, answer the query - not for whether the agent's inputs were solid enough to justify the action it took. This paper is naming a gap that shows up constantly in production agents: tool-calling and retrieval get engineering attention, but "do I actually know enough yet" rarely does.
It's a useful frame, though frameworks like this live or die on whether anyone builds the scaffolding, rather than just citing the paper.