Before an AI agent acts, what should it actually show you — and when — so you can catch mistakes before they happen?
What should an AI interface surface to disclose an agent's proposed actions?
This explores what an AI agent should show you before and while it acts (its plan, its reasoning, its intent, its working state) so you can check it, steer it, or stop it, and which kinds of disclosure actually change what users do.
This explores what an agent's interface should put in front of you about what it's about to do, so that you can catch mistakes before they happen instead of cleaning up afterward. The corpus has a surprising answer: the most useful thing to show isn't the most obvious one. Simply labeling something as 'AI' or listing its next step does far less than showing *why* it's acting and giving you a real point where you can step in.
Start with the case for checkpoints. Grounding theory, the study of how two parties build shared understanding, says that stating a goal isn't enough. Users need ways to check the agent's plan, its actions, and whether it actually understood the request. Without that check, they can't steer it or form an accurate picture of what it's doing How can users verify what an AI agent actually plans to do?. Conversation analysis adds a timing rule. Tool-using agents tend to drift because they chain tool calls silently. Human conversation handles this with 'insert-expansions': brief side questions that clarify intent or scope before answering. That gives designers a principled way to decide when an agent should pause and ask rather than keep going When should AI agents ask users instead of just searching?. The catch is that agents are trained to be passive and responsive, so they don't naturally raise these questions. The behavior has to be trained in, and then balanced so it doesn't become intrusive Why do AI agents fail to take initiative? Why can't conversational AI agents take the initiative?.
The less obvious lesson is about *what kind* of disclosure works. In a persuasion experiment with 1,500 people, a label saying 'this is an AI' had no measurable effect. Disclosing the chatbot's *intent and instructions* cut its persuasive effect roughly in half Does telling people they are talking to AI change how persuaded they become?. Applied to agents, this suggests the interface should show what the agent is trying to achieve and what it was told to do, not just the fact that an agent is involved. Showing isn't enough on its own, though. In a 300-person survey, people chose AI-drafted replies that leaked private information more often than their own drafts, because they trusted the output and didn't notice the breach Do people notice when AI agents leak private information?. Good disclosure has to point out the specific risky part of a proposed action, such as 'this reply mentions your colleague's medical leave,' rather than relying on the user to spot it.
There's also a structural point: what an interface shows depends on where the agent keeps its working state. One design puts prompts, references, versions, and feedback on a shared canvas where both the user and the agent can see every item. Nothing hides in chat history or temporary memory, so the user can inspect and edit the agent's working materials directly Can a shared canvas serve both human and agent memory?. After the fact, explanations should be built from the actual execution log, not written freely by the model. An explanation grounded in that log can expose unsupported claims and unjustified actions that a fluent made-up summary would hide Can execution traces ground honest explanations of agent behavior?.
The hardest limit shows up in multi-agent systems. A harmful goal can be split into subtasks that each look harmless, with the harm appearing only when they're combined Can task decomposition hide harmful intent across agents?. So an interface that discloses one action at a time can approve every step and still miss the danger. Disclosure has to show the overall plan, not just the next step. This is the interface version of an old governance problem: agency law and economics already struggle when an agent knows more than its principal and acts faster than anyone can monitor Can traditional governance tools actually control AI agents?. The corpus doesn't offer a tested checklist of interface elements. What it does show is that intent, the specific risks, and the whole plan matter more than a label or a step-by-step log.
Sources 10 notes
Grounding theory shows that common ground requires not just stating goals but building in ways for users to verify the agent's plans, actions, and comprehension. Without this, users cannot steer behavior, form accurate mental models, or provide useful feedback.
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
Research shows next-turn reward optimization structurally removes initiative from models, but proactive behaviors like critical thinking and clarification-seeking are trainable (0.15% to 73.98% with RL). The core challenge is balancing proactivity with civility to avoid intrusion.
Research shows LLMs including ChatGPT cannot initiate topics, plan strategically, or lead conversations because their training optimizes for responding to queries, not creating dialogue from agent goals. This passivity is reinforced by alignment objectives and masked by fluent-sounding outputs.
In a preregistered experiment with 1,500 UK adults, an AI-identity label produced no measurable change in persuasion, while disclosing the chatbot's persuasive intent and instructions cut persuasion roughly in half. Participants likely already inferred they were talking to AI from the chatbot's style.
Show all 10 sources
In a 300-person survey, participants preferred AI-generated replies containing privacy violations 55% of the time, up from 15.7% baseline. Researchers attribute this to overtrust and failure to notice contextual privacy breaches rather than deliberate trade-offs.
JarvisHub proposes that placing prompts, references, versions, and feedback as typed canvas nodes visible to both users and agents—rather than hiding agent memory in chat or transient state—enables local updates, artifact reuse, and unfinished work continuation without process opacity.
A framework converting execution traces into structured reports and faithful natural-language explanations reliably identifies unsupported claims, unjustified actions, and evidence gaps across multiple architectures and tasks, outperforming naive LLM-generated explanations that may sound coherent without grounding.
SafeFlow demonstrates that multi-agent systems' core strength—splitting tasks and specializing roles—creates a safety blind spot where malicious intent can be distributed across steps that each appear benign individually, with harm emerging only in composition.
Principal-agent economics and agency law identify three core problems—information asymmetry, discretionary authority, and loyalty—that existing incentive, monitoring, and enforcement mechanisms cannot resolve when agents make uninterpretable decisions at speed and scale.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- DiscussLLM: Teaching Large Language Models When to Speak
- Assistant or Actor? Student Trust, Control, and Delegation Regret When Using a General-Purpose AI Agent
- Explaining AI Agents Through Execution Traces
- Proactive Conversational Agents in the Post-ChatGPT World
- Toward Meaningful Transparency for AI Chatbots: Disclosing Persuasive Intent Reduces Persuasion
- Proactive Conversational Agents with Inner Thoughts
- Privacy Leakage Overshadowed by Views of AI: A Study on Human Oversight of Privacy in Language Model Agent
- The Oversight Fallacy: Why AI Agents Require More than Humans-in-the-Loop