AI agents that split planning from doing frequently disagree about what's actually happening in the world.
Researchers studied agent systems built around two separate roles: a planner that decides what to do, and an actor that carries it out, a common setup for tasks like navigating virtual environments over many steps. They had both roles report structured assertions about the task state and compared the reports programmatically, turning up frequent outright contradictions, a problem the paper calls "planner-actor state mismatch." Simply giving agents more task-state information cut down on these mismatches. The team then built a method called Consistent Plan-Act, which feeds detected contradictions back to both roles and fine-tunes them on curated, consistent interactions.
This is a quiet but important finding for anyone building multi-agent AI systems. The planner and the component executing its plan can hold different internal pictures of reality, and that gap silently drags down performance even when each piece looks fine on its own. In one tested environment, the fix lifted task success from under 39% to over 54%, a reminder that coordination, not raw model capability, is often the hard part of agentic AI.
It's the AI equivalent of a manager and a line worker reading from two different memos. The fix here is retraining both sides to agree on the memo, not just asking them to talk more.