Microsoft's AI design rules were built for suggestion boxes — do they still hold up for AI agents that act on their own?
How do Amershi's AI interaction guidelines compare to agent-specific UX frameworks?
This explores how Microsoft's 18 general-purpose human-AI interaction guidelines (Amershi et al.) hold up next to newer design frameworks built specifically for AI agents that plan, act, and use tools on a user's behalf.
This explores how Microsoft's 18 general-purpose guidelines for human-AI interaction hold up next to newer design frameworks built for AI agents that plan and act for you. The short version is that Amershi's guidelines are organized around time: what the system should do at first contact, during the interaction, when it gets something wrong, and over time How should AI systems be designed across different interaction phases?. Agent frameworks are organized around something else: who decides, who acts, and when control passes between the human and the agent. The two overlap, but they answer different questions.
Amershi's guidelines came from more than 150 design recommendations and were checked with 49 practitioners across 20 products. Most of those products were features like recommendations, autocomplete, and ranking, where the AI suggests and the user decides. Agent-specific work begins where that assumption fails. Magentic-UI, for example, names six mechanisms: co-planning, co-tasking, action guards, verification, memory, and multitasking. They exist because there is no reliable answer to when an agent should stop and ask a human When should human-agent systems ask for human help?. Amershi's "when wrong" phase assumes you can notice a mistake and fix it afterward. An agent that has already sent the email or booked the flight needs guards before it acts, not just repair afterward. Cocoa goes a step further. It lets users assign each step of a plan to themselves or to the agent and revise the plan while it runs. In a lab study this gave users more control than a chat interface without making the system harder to use Can users steer AI agents better by revising plans mid-execution?. That kind of shared, divided labor has no real counterpart in the 2019 guidelines.
A second gap concerns initiative. Amershi's guidelines mostly assume the AI waits until the user does something. Agent research shows that waiting is itself a design problem. Models are passive because of how they're trained, not because they can't do more: optimizing for the next reply removes initiative, but training can bring behaviors like asking clarifying questions back Why do AI agents fail to take initiative? Why can't conversational AI agents take the initiative?. Conversation analysis offers a concrete rule for when an agent should pause and ask instead of quietly chaining tool calls. It borrows "insert-expansions", the side questions people ask mid-conversation to clarify what someone wants before answering When should AI agents ask users instead of just searching?. That turns Amershi's broad advice to "scope services when in doubt" into a specific mechanism for agents.
A deeper challenge is that agent frameworks question whether "interface design" is still the right unit. One argument is that AI context, meaning the prompt, history, retrieved data, and hidden state, keeps changing in ways users can't learn the way they learn a fixed UI. That calls for context engineering more than interface design How does AI context differ from conventional software context?. AXIS goes further: agents that call APIs directly instead of clicking through UIs finished tasks 65–70% faster and reduced users' cognitive workload Can API-first agents outperform UI-based agent interaction?. In some cases the best agent experience means less interface, not a better one. The scope widens beyond the single user, too. Some researchers argue agents must balance many stakeholders' interests with visibility and accountability, not loyalty to one user Should AI agents serve one user or balance many interests?. Executives are split on whether agents are tools or coworkers Does agentic AI fit the tool-or-coworker framework?. Amershi's framework treats AI as a tool throughout.
One caveat about this corpus: it has no paper that directly compares Amershi's guidelines against an agent framework item by item, so this comparison is assembled from adjacent work. What it does suggest is that Amershi's guidelines still work as a minimum standard, since agents still need to explain their capabilities, support correction, and learn over time. The agent frameworks add a new kind of design problem: deciding who holds control at each step, before any action happens. A timeline-based framework doesn't capture that.
Sources 10 notes
Microsoft researchers synthesized over 150 design recommendations into 18 actionable guidelines grouped by when they apply: initially, during interaction, when wrong, and over time. Validation with 49 practitioners across 20 products confirmed their applicability.
Magentic-UI identifies co-planning, co-tasking, action guards, verification, memory, and multitasking as mechanisms that work around the lack of ground truth for optimal deferral timing. Rather than solving the timing problem directly, these mechanisms distribute decision-making across multiple touchpoints.
Cocoa's system, which lets users assign plan steps to themselves or an agent and revise plans mid-execution, afforded greater steerability than chat-baseline interaction in a lab study (n=16) while maintaining ease of use, with field users valuing explicit work delegation.
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.
Show all 10 sources
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.
AI interactions operate on a substrate of constantly shifting context—prompt, history, retrieved data, hidden state—that users cannot internalize like traditional UIs. This structural mutability demands a new design discipline centered on context engineering rather than interface design.
The AXIS framework shows that prioritizing API calls over sequential UI interactions cuts task completion time by 65–70% while maintaining 97–98% accuracy and reducing cognitive workload by 38–53%. A self-exploration mechanism automatically discovers and constructs APIs from existing applications, solving the bootstrapping problem.
Research argues the traditional single-agent-single-user alignment framing is insufficient at scale. Instead, governance should rest on inclusivity (balancing diverse interests), visibility (rigorous tracking and transparency), and liability (developer and deployer accountability for harms).
A survey of 2,102 executives across 116 countries found 76% view agentic AI as a coworker rather than a tool. This duality creates operational tensions because agentic systems simultaneously substitute and complement labor, requiring both HR and asset management approaches that traditional either-or frameworks cannot accommodate.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants
- DiscussLLM: Teaching Large Language Models When to Speak
- Assistance or Disruption? Exploring and Evaluating the Design and Trade-offs of Proactive AI Programming Support
- Proactive Conversational Agents in the Post-ChatGPT World
- Proactive Conversational Agents with Inner Thoughts
- Cocoa: Co-Planning and Co-Execution with AI Agents
- Guidelines for Human-AI Interaction
- Towards a Science of Scaling Agent Systems