A new paper argues that how AI agents communicate is just as important as what they can do - and that most agent research has overlooked it entirely.
Researchers formalized what they call "Communication Policy" - the rules governing when and how a large language model agent reaches out to users versus acting on its own. They tested two interaction channels, text-based and structured UI, across varied task environments, user personas, and model combinations, using two experimental setups: one modeling a single user-agent relationship, another modeling a planner-executor architecture. The results show the channels have complementary strengths: text interaction tends to improve raw task completion rates, while structured UI leads to better response quality and closer adherence to a persona. A hybrid approach combining both outperformed either channel in isolation.
The more interesting claim is Communication Policy Evolution (CPE), a self-improvement loop that refines these policies through prompt adjustments alone - no model modification required. That matters because the dominant assumption in agent development is that better behavior requires better models. Finding that prompt-level refinement can close measurable gaps in task success is a useful data point for teams constrained by compute or locked to a particular model tier.
Whether prompt-evolved communication policies hold up outside controlled lab settings remains to be seen - but treating communication itself as a design variable with real costs is a framing the field has been slow to adopt.