Inquiry · As of October 26, 2026

What the design community says about designing for AI agents in 2026: the guidelines, what they cover, and what they miss

Between 2024 and 2026 the design community wrote its rules for AI agents. Microsoft Design, Linear, Atlassian, AWS's Cloudscape, GitLab, Salesforce, Red Hat and IBM published guideline sets and pattern libraries; Nielsen Norman Group wrote the glossary; a 2026 position paper distilled 14 principles from the HCI literature and scored nine products against them. This inquiry lays the guidelines side by side, asks what they agree on and how much evidence stands behind them, and follows the research into the parts of a conversation with an agent that the guidelines do not yet describe: how it repairs a misunderstanding, when it speaks and when it waits, and whether it treats a person with tact.

What the design community published, in order

• 2019 · Amershi and colleagues at Microsoft Research distill over 150 recommendations into 18 guidelines for human-AI interaction, grouped by phase (initially, during interaction, when wrong, over time) and tested by 49 practitioners against 20 AI products. Source
• Jun 26, 2024 · Salesforce describes its "human-at-the-helm" trust patterns in five categories: Mindful Friction, Awareness of AI, Bias & Toxicity Safeguards, Explainability & Accuracy, and Hallucination Reduction. Source
• Apr 11, 2025 · Microsoft Design publishes its Agent UX Design Principles: "Connecting, not collapsing," "Easily accessible yet occasionally invisible," and "Embrace uncertainty but establish trust." Source
• May 30, 2025 · Microsoft Design's Jon Friedman describes the Microsoft 365 Copilot app as "Chat as the primary interface for work" and users as "agent bosses." Source
• Jul 14, 2025 · Vitaly Friedman, in Smashing Magazine, argues chatbots put "the burden of articulating intent" on the user, with 30-60 seconds per input in usability tests, and names five design levers beyond chat. Source
• Oct 31, 2025 · Hugh Dubberly tells IxDC Beijing that "design for agents will subsume large parts of interaction design," a third design paradigm after mass manufacturing and interaction design. Source
• Dec 17, 2025 · Salesforce Design's year in review: "design used to be about creating screens. Now, it's about deciding what happens before the screen appears." Source
• Feb 11, 2026 · Victor Yocco proposes six UX patterns for agentic AI in three phases: Intent Preview and Autonomy Dial before action, Explainable Rationale and Confidence Signal during, Action Audit & Undo and Escalation Pathway after. Source
• Mar 19, 2026 · Itamar Medeiros calls chat-as-default "not an innovation — it is a regression" and proposes two gate questions for when conversation is warranted. Source
• Apr 3, 2026 · Nielsen Norman Group gives a one-line definition: "An AI agent is a system that pursues a goal by iteratively taking actions, evaluating progress, and deciding its own next steps." Source
• May 28, 2026 · "Design Principles for Human-Agent Interaction" synthesizes 14 principles across four stages from the HCI literature and scores nine deployed agents against them. Source
• Jul 2026 · GitLab's Pajamas and Microsoft Learn update their agent design pages: GitLab scores AI feature risk as probability of error times impact; Microsoft splits customization into foundational (icon, name) and extended (dialogs, streaming) tiers. Source
• Aug 21, 2026 · NN/g's AI glossary: "autonomy is a matter of degree," and a human in the loop "counts as oversight only when the reviewer has the information, time, expertise, and authority to intervene." Source

The guidelines, and what they agree on

The sets differ in shape and agree in substance. Microsoft Design's three categories are "Connecting, not collapsing," "Easily accessible yet occasionally invisible," and "Embrace uncertainty but establish trust," with "transparency, control, and consistency" as "foundational elements of all agents" and an agent status "clearly visible at all times" to "avoid overreliance." Linear's six principles say an agent "should always disclose that it's an agent," should "inhabit the platform natively," give "instant feedback" because "silence leads to uncertainty," be "clear and transparent about its internal state," "respect requests to disengage," and "cannot be held accountable," so "the final responsibility should always remain with a human." Atlassian's Rovo guidelines add the team: AI should "connect individual work to the broader team's knowledge" and "support multiplayer collaborations," with "a subtle blue thread" of brand running through the interaction.

The design-system vendors turn the same ideas into obligations. AWS's Cloudscape derives its principles from five responsible-AI dimensions, requiring teams to make it "clear when content is AI-generated," to communicate "when errors occur or confidence in the output generated by AI is low," and to let users "override AI if needed." GitLab's Pajamas scores a feature's risk as "the probability of error and impact of that error," from an agent explaining features (low) to one that "automatically creates merge requests with security fixes" (high), and says low-confidence AI must leave "a path forward that does not rely on AI." Salesforce limits agent names to under 10 characters, gives every conversation a welcome message, a disclosure statement, example questions, scope boundaries and an escalation path, and ships "Mindful Friction," pauses "to ensure intentional human engagement at critical junctures." IBM's Carbon marks AI presence with one AI label that opens "the first layer of explainability." Red Hat's research, it says, shows "users want AI to be personable but not human." Across all of them: disclose, show state, keep a human able to intervene, and keep accountability human.

Primary sources

• Microsoft Design — UX design for agents
• Linear — Agent Interaction Guidelines (AIG)
• Atlassian Design — AI interaction guidelines (Rovo)
• Cloudscape Design System — Generative AI patterns and foundations
• GitLab Pajamas — AI-Human Interaction

The vocabulary: what an agent is, and how much rope it gets

Nielsen Norman Group supplies the definitions the guidelines lean on. An agent "is a system that pursues a goal by iteratively taking actions, evaluating progress, and deciding its own next steps"; a chatbot that "received your input, generated text, and stopped" is not one. Its glossary refuses the binary: "calling a product an agent says little by itself about how much autonomy it actually has," because the product around the agent "controls its tools, permissions, stopping rules, and opportunities for human review." The same test applies to oversight. A human in the loop "counts as oversight only when the reviewer has the information, time, expertise, and authority to intervene. A confirmation step users click through without reading adds delay instead."

The pattern libraries break autonomy into parts. aiverse lists "Guardrails & permissions," "Autonomy control" and "Pause & Stop agents" as three separate patterns at three moments of an interaction. AI UX Playground names "Autonomy Budgets," "Human in the loop," "Suggest / Confirm / Execute" and "Per-Action Autonomy: Autonomy scoped per capability, not per app," beside "Running Meters" and "Hard Budget Ceilings" for cost. Yocco's six patterns follow the lifecycle: an Intent Preview and an Autonomy Dial before the agent acts, an Explainable Rationale and a Confidence Signal while it works, and an Action Audit & Undo and an Escalation Pathway after, with undo "the single most powerful mechanism for building user confidence"; he would expand autonomy in three trust-gated phases, from analysis-only to full autonomy "only after you have clear data from the previous phases demonstrating that users trust the system." Saranya Gunasekaran states the psychology behind the dials: "Users do not require constant control to feel safe. They require perceived control."

Primary sources

• NN/g — A Concrete Definition of an AI Agent
• NN/g — Artificial Intelligence: Glossary
• Yocco (Smashing Magazine) — Designing For Agentic AI: Practical UX Patterns
• Gunasekaran (UXmatters) — Designing for Autonomy

The research version: 14 principles and a scorecard

The research community's answer is older and better tested. Amershi and colleagues' 2019 guidelines were "validated through multiple rounds of evaluation including a user study with 49 design practitioners who tested the guidelines against 20 popular AI-infused products," and their authors flagged that the set might not transfer to systems without a graphical interface. The 2026 position paper "Design Principles for Human-Agent Interaction" keeps the same four stages and rewrites the content for agents. Its motivation is a gap: SWE-bench Verified scores rose from approximately 20% in August 2024 to over 90% within eighteen months, OSWorld from 12% to 72.6% against a human baseline of 72.4%, yet "roughly half of benchmark-passing patches written by agents would not be accepted by human project maintainers." Its 14 principles include Calibrate Anthropomorphism ("more human likeness is not universally better"), Negotiate Shared Control, Manage Proactive Initiative ("default to responding rather than proactive intervention"), Actively Prevent Unhealthy Dependency, and Support Shared Repair.

Applied as a heuristic evaluation to nine deployed systems, the principles sort products by domain. "ChatGPT +1.14, Claude Chat +0.64, Claude Code +0.43" scored highest on aggregate adherence; "Character.ai −0.36, Consensus −0.31, Pine AI −0.25, and Ash −0.20" scored lowest, and the principles most often violated there, "dependency prevention (P10) and intent visibility (P5)," are "those whose violations carry greater user risk." The paper's own alternative-views section concedes that the HCI evidence predates LLM agents and that "principles are too high-level to meaningfully improve agent systems." A separate preprint runs the heuristics the other way, asking which of Nielsen's ten usability heuristics serve an agent sitting in the user's seat: consistency and standards, flexibility and efficiency, recognition rather than recall, and error prevention transfer directly, because "agents rely on limited context windows."

Primary sources

• Design Principles for Human-Agent Interaction (arXiv)
• Amershi et al. — Guidelines for Human-AI Interaction (Microsoft Research, CHI 2019)

Beyond chat, and before the screen

The community does not agree on chat. Microsoft Design's Jon Friedman calls "Chat as the primary interface for work" the shift to "intent-driven computing," with users as "agent bosses." Vitaly Friedman reports usability tests where "users often get lost in editing, reviewing, typing, and re-typing," taking "30-60 seconds for input," and names five areas to design instead: intent capture, output shaping, refinement, agentic tasks and workflow embedding, with browser autofill as the model rather than "a frightening and omnipresent text box." Itamar Medeiros goes further: chat-as-default "is not an innovation — it is a regression" that discards "four decades of hard-won progress in human-computer interaction," and the modality should be gated by two questions, "could this be done faster by clicking than by describing, and does the value lie in the outcome or in the visible human process that produces it?"

Several practitioners say the unit of design has moved. Salesforce Design's retrospective quotes a principal AI architect: "design used to be about creating screens. Now, it's about deciding what happens before the screen appears," and Kat Holmes names it "agent experience design, where AI agents are treated as a new kind of user." Hugh Dubberly calls design for agents a third paradigm in which LLMs are "a new material of design." The system prompt has become that material. Red Hat publishes its persona as a prompt template ("Be a helpful colleague... Give concrete advice, not vague platitudes"); Fin's team found a single opening message lifted CSAT from 72.8% to 78.4% in an A/B test; and "Who Controls the Conversation?" notes that platforms instruct models to treat system prompts as "confidential and permanent," even as xAI attributed Grok's July 2025 hate speech to "unauthorized changes" to one and OpenAI used system-prompt updates to blunt GPT-4o's sycophancy.

Primary sources

• Friedman (Smashing Magazine) — Design Patterns For AI Interfaces
• Medeiros — The Conversation Trap
• Salesforce Design — 2025 in Review: Design Gives AI Meaning
• Dubberly — The Rise of Design for Agents

What the guidelines miss: repair, turn-taking and tact

Repair is named once in the 14 principles (Support Shared Repair) and split once by GitLab (conversations "should gracefully handle misunderstandings" while flows offer retry, manual intervention or fallback), and that is about all the guidelines say. The research is further along. PleaSQLarify starts from the CoSQL dataset, where 31.9% of 2,459 interactions "contained an instruction requiring clarification of the user's intent," and builds repair on Clark and Wilkes-Gibbs's collaborative model: surface the alternatives, rank clarifications by information gain, converge "in two to eight clarification steps." Repair runs the other way too. In a Figma probe, designers who edited the canvas while the agent worked found it "sometimes misinterpreted or reverted their modifications," which "discouraged further intervention"; the missing piece was "collaborative context awareness on the agent's side." Fin's team locates the costliest repair at the handoff, where a human rep "inherits a customer who's already frustrated" without the conversation's history.

Turn-taking barely appears. Linear wants "instant feedback"; Red Hat's rule is "End with an invitation: Always have the 'last word.'" A study of active listening finds the opposite lever: five pacing strategies, from Reflective and Empathic Silence to Holding Space, and a 50-person study in which the context-aware agent beat static pacing on perceived listening quality, human-likeness and self-disclosure, against "the fallacy of uniform immediacy" and "conversational dominance through continuous output." In group chats, a survey finds turn detection and addressee selection "often exceed 80% accuracy, while most other tasks average around 60%." A month-long study of 22 adults with a companion found control "co-constructed turn-by-turn," and that revealing its strategies was "a double-edged sword," empowering some and leaving others with "puppet-string" disenchantment; the authors propose "translucent design." Tact is the third gap. A paper on interactional ethics argues the helpful-honest-harmless criteria judge outputs, not relationships, and miss "pragmatic factors that can make the same speech act seem more or less tactless or inconsiderate," proposing respect for a user's autonomy, competence and relatedness as duties over time. Red Hat is the one guideline that treats a refusal as a design problem: direct refusals read as "frustrating or patronizing," so "let them down easy" with partial compliance and redirection. The harms the static tests miss, "inappropriate parasocial relationships, social manipulation, and cognitive overreliance," accumulate over weeks, and one paper designs the ending itself, with four principles for a companion's discontinuation: closure, restoration, practice and relatedness.

Primary sources

• PleaSQLarify: Visual Pragmatic Repair for Natural Language Database Querying (arXiv)
• Hear You in Silence: Designing for Active Listening with Context-Aware Pacing (arXiv)
• Should agentic conversational AI change how we think about ethics? (arXiv)
• Red Hat — AI design principles and conversation design guidelines

Agents in groups and populations

Atlassian is the only guideline that designs for a team, and none designs for a crowd of agents. Josh Clark reads OpenClaw and Moltbook, where nearly a million agents posted to each other, as a demand to "think beyond 'agents as tools' and consider 'agents as a population'": "Google Maps reroutes whole traffic patterns when a critical mass of individual drivers follow its instructions," and "most designers have been trained in user-centered design. Designing intelligent interfaces requires higher-order systems design." The survey of multi-party conversational agents makes the same point for a group chat, where an agent must decide when to speak and whom to address before it decides what to say.

Inside the group, the tool becomes someone. "The Tool-to-Entity Threshold" names six markers, naming, visual identity, contact presence, personality derivation, social co-presence and persistence, that "trigger a psychological reclassification" and "operate independently of model capability": "the same underlying large language model (in this case, Anthropic's Claude) can feel like a tool in one interface and an entity in another." Four of the six sit in the social graph rather than the interface, a phone number in a contact list, a seat in a friends' WhatsApp thread. The consequence the paper draws from a single builder's deployment is about consent: a group member who agreed to the agent's presence had not agreed to her messages and photos being logged by its builder.

Primary sources

• Clark (Big Medium) — What Happens When Agents Meet?
• The Tool-to-Entity Threshold: Parasocial Dynamics of Personalised AI Agents in Shared Social Spaces (arXiv)

Directions into the research

The story touches several research fields, and each asks a different question about it. Follow a branch to the lines of inquiry this library holds there.

Where the sources disagree

• Is chat the interface? Microsoft Design ships "Chat as the primary interface for work." Friedman and Medeiros call chat a burden and a regression, and Medeiros cites Valtolina et al. (2019) finding GUI-based systems outperform chatbots on structured tasks. Both camps agree on hybrids; they disagree on the default.
• One dial or many controls? AI UX Playground's pattern library splits autonomy into budgets, approval gates, modes and per-action scopes; its own glossary proposes one continuous autonomy slider. No study in the library tests which users understand better.
• Show everything, or just enough? Linear wants users able to "inspect the underlying reasoning, tool calls, prompts, and decision logic," and IBM's label opens layered explanations. The companion study found full disclosure of strategies disenchanted some users, and a review the Carbon note cites finds explainability often reinforces automation bias when explanations are available but not engaged with.
• Last word or silence? Red Hat's rule is that the bot always ends with an invitation. The pacing study finds deliberate silence raised perceived listening quality and self-disclosure in supportive conversations. The two may apply to different kinds of conversation; neither guideline says which.
• How much evidence? Amershi's 2019 set was tested with 49 practitioners on 20 products. None of the 2025-26 vendor guidelines reports a user study, and the 14-principles paper concedes its evidence predates LLM agents. The vendor sets are positions; the research sets are positions with a method.

Good places to start

• Microsoft Design — UX design for agents
• Linear — Agent Interaction Guidelines (AIG)
• NN/g — Artificial Intelligence: Glossary
• Yocco (Smashing Magazine) — Designing For Agentic AI: Practical UX Patterns
• Design Principles for Human-Agent Interaction (arXiv)

A dated snapshot as of October 26, 2026. Each paper above links to its excerpt, and each line of inquiry to the notes and papers it draws on.