Abstract
As AI agents evolve from simple chatbots into autonomous systems capable of browsing the web, executing code, and managing digital workflows, the "communication gap" between humans and machines is widening. This paper identifies 12 critical challenges in human-agent communication—ranging from how agents should explain their plans to how they should learn from feedback—and provides a roadmap for building more transparent, controllable, and trustworthy AI collaborators.
Introduction
We are entering the era of "agentic" AI. Unlike traditional AI models that simply classify images or generate text, modern agents (powered by Large Language Models like GPT-4) can perceive their environment, use tools, and take actions to solve complex, multi-step problems.
However, this increased power comes with increased risk. An agent tasked with a high-level goal might take unintended actions, such as signing a user up for a new credit card to secure a flight deal or accidentally deleting critical files during a code fix. The core problem is a lack of common ground: a shared understanding between the human and the agent regarding goals, capabilities, and the state of the world. Without effective two-way communication to establish this ground, users cannot effectively monitor, verify, or steer their AI agents.
Framework and Methodology
This work synthesizes decades of research from psychology, cognitive science, and Human-Computer Interaction (HCI)—specifically drawing on Clark & Brennan’s "grounding in communication" theory—to analyze the unique hurdles posed by modern generative foundation models. By drawing on experiences building complex agentic systems like Microsoft’s AutoGen and Magentic-One, the authors categorize the communication landscape into 12 distinct challenges mapped across three interaction phases: Before, During, and After execution.
Key Insights into Agentic Interaction
The Shift from Classifiers to Agents
Traditional AI was often a "one-shot" interaction (e.g., "Is this a cat?"). Agents involve long-horizon, multi-step workflows where the "process" is just as important as the "result." This shift necessitates a move from simple output verification to continuous process monitoring.
The Stochasticity Paradox
Even with deterministic settings, agents operating in dynamic environments (like the live web) can produce different sequences of actions for the same prompt. This makes it incredibly difficult for a user to build a reliable "mental model" of how the agent behaves (Challenge X2).
The "Level of Detail" Tightrope
There is a fundamental tension between transparency and cognitive load. If an agent reports every single sub-step, the user is overwhelmed (Challenge X3); if it reports nothing, the user loses control. Achieving the right balance is critical for usability.
Operational Alignment vs. Cognitive Similarity
The authors clarify that "common ground" in AI doesn't mean the agent "understands" like a human. Instead, it refers to operational alignment—ensuring both parties have the same inferences about the current state and future outcomes.
Taxonomy of 12 Challenges
The paper’s primary contribution is a structured taxonomy of communication hurdles:
- General Challenges (X1-X4): These include the verification of behavior, maintaining consistency across runs, choosing the right level of detail, and managing long-term interaction context.
- User to Agent (U1-U3): These focus on communicating high-level goals, specifying nuanced preferences (e.g., "minimize cost vs. minimize time"), and providing feedback for future improvement.
- Agent to User (A1-A5): These involve disclosing capabilities/limitations, explaining intended plans, providing real-time status updates, reporting environmental side effects (Challenge A4), and confirming goal achievement.
Future Directions
The authors issue an "urgent call to action" for AI researchers and HCI designers to collaborate on building systems that prioritize "common ground" as a first-class citizen in agent architecture. Key areas for future work include:
- New Design Patterns: Moving beyond "chat bubbles" to more sophisticated interfaces for plan inspection and state visualization.
- Agent "Cards": Developing dynamic "agent cards" (similar to model cards) that show what an agent can and cannot do in a specific context.
- Addressing Side Effects: Developing better ways for agents to communicate "latent" changes to the environment that aren't immediately obvious in a chat interface.
Summary updated