Shuichiro Ogawa
日本語

Notes · updated 2026-06-27

Industry Discourse on Agentic Experience (AX): What Vendors, Consultants, and Practitioners Are Saying

By the first half of 2026, “Agentic Experience (AX)” — in which AI agents autonomously execute tasks — had become a central topic across industry. John Maeda’s Design in Tech Report 2026 proposed the shift “from UX to AX”1, and that same year, companies unveiled agent-oriented features at Figma Config 2026, Adobe MAX, Canva Create, and WWDC26.

Yet the term AX lacks a unified definition. Different companies and thought leaders frame and implement AX in divergent ways. This note draws on 73 industry sources (24 T2 analyst reports, 18 T1v vendor publications, 16 T3 expert-opinion pieces, and 15 T3 developer-voice accounts) to map the landscape of AX discourse in industry2.

Related: agentic-experience-literature (academic foundations) / agentic-experience-design-synthesis (academic-industry comparison) / design-agent-tools-landscape-2026 (tool landscape) / mcp-design-agent-integration (MCP integration) / the roundtable on delivering multi-agent value (value of multi-agent systems).

The Definition of “From UX to AX” Is Fragmented

Four distinct lineages characterize how major voices define AX.

Maeda’s “decision surface” thesis. Maeda defines AX as the transition from the screen as a “work surface” to a “decision surface”3. Where UX asks “how do I operate this?”, AX asks “can I simply state what I want and have it delivered?” He argues that agents should be designed as action-outcome-feedback-correction loops, and identifies three layers of AX: Capabilities (what the agent can do), Trust (how control is maintained), and Protocols (how systems connect). He also positions AX as the realization of the tenth law from his “Laws of Simplicity” — “subtract the obvious and add the meaningful”4.

Nielsen’s “Intent Layer” thesis. Jakob Nielsen argues not for the death of UX but for its metamorphosis5. He frames the transition as moving from Conversational UI (a passive tool waiting for prompts) to Delegative UI (an agent that autonomously plans, executes, and iterates), and proposes the “Intent Layer” as a new design object6. The Intent Layer is a mission definition that includes goals, budgets, and constraints — rewriting the very definition of usability. The design focus shifts to policy surfaces (permissions, cost ceilings, ethical boundaries), trust conveyors (provenance chips, uncertainty meters, rollback mechanisms), and system “temperament” design (proactive versus passive).

Salesforce’s “bidirectional design” thesis. Salesforce officially defines AX as “the development and optimization of digital environments so that AI agents can operate efficiently and orchestrate human-centered outcomes”7. What distinguishes this definition is its bidirectionality: it encompasses both “designing for agents” and “designing by agents,” reframing the designer’s role from “interface architect” to “experience orchestrator.” The company identifies three design requirements: Intent-first architecture, Cross-platform agent orchestration, and Real-time capability discovery8.

Apple and Google’s “declarative primitives” approach. Apple designed agentic experiences not as UI navigation but as direct invocation of structured intents, through App Intent Domains (WWDC25) and Dynamic Profiles (WWDC26)9. Google chose a design in which agents define the UI through declarative JSON via the A2UI (Agent-to-UI) protocol10. The open-sourcing of DESIGN.md — a markdown file that describes brand rules in YAML — represents an attempt to transform design systems into persistent agent context11. Both companies embody AX through the technical approach of “defining experiences in data, not executable code.”

NNg does not adopt the term AX directly, instead offering the framing that “AI agents should be included as users in the design scope” alongside the concept of “Context Architecture” — applying the principles of information architecture to AI1213.

Convergence of Design Principles, Divergence of Intervention Granularity

Although definitions are fragmented, design principles converge on “transparency, control, intent-first, and preservation of human judgment.”

Microsoft articulates three principles — “transparency, control, and consistency” — and defines the agent as assisting rather than replacing user judgment14. Anthropic advances five principles: “maintaining human control, transparency, value alignment, privacy, and inter-agent security”15. OpenAI makes “the user is always in control” its foundational principle, designing for permission requests before action and user-initiated interruption or stoppage at any time16. Sheta Chatterjee (Head of UX, Figma) states that “the user’s focus should always be on the goal, never on managing AI”17. Adobe codifies a division of labor: “creators own vision, judgment, and direction; agents own orchestration and execution”18.

The principles converge, but the granularity of intervention differs across companies. Harrison Chase (LangChain) uses the concept of “leash length,” positioning v0 as short leash (returns quickly), Cursor as middle, and Devin as long leash19. This concept corresponds to the academic finding on “optimal confirmation frequency” by Zhou et al. (2026), discussed in agentic-experience-literature.

The Gap Between Hype and Reality: Quantitative Evidence

Analyst research quantitatively documents the gap between AX hype and reality.

Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 202720. It simultaneously warns of “agent-washing” — rebranding existing products without substantive agent functionality — estimating that of thousands of vendors, only approximately 130 offer genuine capabilities21.

Forrester, in its mid-2026 assessment, finds that the majority of “agentic” experiences remain conversational, with true autonomy rare, and hype outpacing consumer behavior22. It also predicts that one-third of enterprises will deploy agents prematurely, damaging brand and CX23.

McKinsey’s survey (1,993 respondents, 105 countries) finds that 62% are experimenting with AI agents, but only approximately 6% qualify as “high performers” who achieve over 5% EBIT impact24. Roughly two-thirds of organizations remain at the experimentation or pilot stage.

Deloitte’s survey (3,235 senior leaders, 24 countries) reports that 23% use agentic AI at a moderate level or above, but only 21% have mature governance models25. Governance is not keeping pace with scaling.

Additional data from Forrester is telling. While 78% of AI decision-makers say they trust AI outputs, only 15% report EBITDA improvement over the past 12 months26. Trust is high, but outcomes have not materialized.

Structural Constraints Reported by Developers

First-hand developer testimony reveals the structural constraints behind the design principles.

Approval fatigue and the hollowing out of oversight. Anthropic’s measured data shows that users approve permission prompts 93% of the time, leading the company to conclude that per-action approval fails to function as genuine oversight. They shifted to a model in which the agent “operates freely within pre-defined boundaries”27. Operating under the recognition that “model-layer defenses will never be 100% effective,” they introduced VM-based isolation — which, however, creates structural friction with enterprise endpoint detection27.

The cost structure of multi-agent systems. Anthropic’s measurements show that multi-agent systems consume approximately 15x the tokens of chat-based interactions28. Harness costs (the supervisory and iterative structure governing the agent) can reach 20x those of standalone generation ($200 vs. $9)29. Early failure cases have been publicly documented, including “spawning 50 sub-agents for a simple query” and “preferring SEO articles over authoritative sources.”

Context as the deciding factor. Harrison Chase states that “success and failure alike come down to context,” arguing that now that model capability has crossed a threshold, the differentiator lies in input context design30. Thomas Lowry of Figma likewise testifies that “when file structure and layer names are unclear, AI misinterprets the design and creates downstream problems”31. Anthropic’s Applied AI team recognizes as a structural constraint that merely expanding the context window does not solve the problem — attention pressure and context contamination occur regardless of window size32.

The limits of self-evaluation. Prithvi Rajasekaran of Anthropic reports that “when you have an agent evaluate its own output, it will confidently praise the result even when quality is clearly low,” establishing the separation of generator and evaluator roles as a design imperative29.

The “Floor and Ceiling” Metaphor

Dylan Field (Figma CEO) remarked at Config 2026 that “AI has lowered the floor but hasn’t raised the ceiling,” a formulation widely cited across industry33. Andrej Karpathy articulated the same structure differently: “vibe coding raises the floor; agentic engineering raises the ceiling”34. Karpathy refers to agents as “interns,” insisting that humans retain responsibility for aesthetics, judgment, taste, and oversight.

Simon Willison has proposed the “lethal trifecta” — a security risk model in which data exfiltration becomes possible when three factors coincide: access to private data, exposure to untrusted content, and the ability to communicate externally — thereby identifying a structural constraint on agent design35. Willison publishes “Agentic Engineering Patterns” as a living document, positioning agentic engineering as a specialized technical discipline36.

What Industry Is Not Talking About

Surveying the 73 industry sources, several conspicuous silences emerge.

  1. Learning pathways for juniors: Vendors celebrate “lowering the floor” (reducing barriers to entry), yet none address how juniors will learn design judgment once that floor has been lowered. The academic literature (agentic-experience-literature, specifically Li et al. 2024 and Luo et al. 2025) provides empirical evidence of this risk.

  1. Structural causes of agent-washing: Gartner warns of “washing” but offers no structural analysis of why it occurs — specifically, the incentive within existing SaaS business models to claim agentic capabilities.

  2. The structure of cost pass-through: Figma has disclosed that its gross margin dropped from 92% to 86% (see design-agent-tools-landscape-2026), yet industry discourse on sustainable models for passing through inference costs remains sparse.

References

Footnotes

  1. Maeda, J. 2026. “Design in Tech Report 2026: From UX to AX.” https://johnmaeda.medium.com/design-in-tech-report-2026-from-ux-to-ax-f9d83164f4d2 ↩

  2. The full corpus is recorded in source/review/design-agent-academic/industry.md. ↩

  3. Maeda, J. 2026. “What is AX?” https://maeda.pm/2026/06/11/what-is-ax/ ↩

  4. Maeda, J. 2025. “Simplicity and Agentic Experience (AX).” https://johnmaeda.medium.com/simplicity-and-agentic-experience-ax-0087553b73d8 ↩

  5. Nielsen, J. 2025. “No More User Interface?” https://jakobnielsenphd.substack.com/p/no-more-ui ↩

  6. Nielsen, J. 2026. “Intent by Discovery: Designing the AI User Experience.” https://jakobnielsenphd.substack.com/p/intent-ux ↩

  7. Holmes, K. 2025. “Welcome to a New Era of Agent Experience Design.” Salesforce Blog. https://www.salesforce.com/blog/agent-experience-design/ ↩

  8. Salesforce. “How to Embrace the Great UX Paradigm Shift to Agentic Experience Design.” https://www.salesforce.com/blog/ux-shift-to-agentic-experience-design/ ↩

  9. Apple. WWDC25 Session 244 “Get to know App Intents.” https://developer.apple.com/videos/play/wwdc2025/244/ ↩

  10. Google. “Introducing A2UI.” https://developers.googleblog.com/introducing-a2ui-an-open-project-for-agent-driven-interfaces/ ↩

  11. Google Labs. “Stitch’s DESIGN.md format is now open-source.” https://blog.google/innovation-and-ai/models-and-research/google-labs/stitch-design-md/ ↩

  12. NNg. “AI Agents as Users.” https://www.nngroup.com/articles/ai-agents-as-users/ ↩

  13. NNg. “Context Architecture.” https://www.nngroup.com/articles/context-architecture/ ↩

  14. Microsoft Design. “UX design for agents.” https://microsoft.design/articles/ux-design-for-agents/ ↩

  15. Anthropic. “Our framework for developing safe and trustworthy agents.” https://www.anthropic.com/news/our-framework-for-developing-safe-and-trustworthy-agents ↩

  16. OpenAI. “Introducing ChatGPT agent.” https://openai.com/index/introducing-chatgpt-agent/ ↩

  17. Figma. “How to Design Agentic Tools for Work.” https://www.figma.com/blog/how-to-design-agentic-tools-for-work/ ↩

  18. Adobe. “The age of creative agents.” https://blog.adobe.com/en/publish/2026/04/15/the-age-of-creative-agents-rise-creative-director ↩

  19. Chase, H. 2025-2026. Sequoia Capital Podcast. https://sequoiacap.com/podcast/context-engineering-our-way-to-long-horizon-agents-langchains-harrison-chase/ ↩

  20. Gartner. 2025-06-25. “Gartner Predicts Over 40 Percent of Agentic AI Projects Will Be Canceled by End of 2027.” https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027 ↩

  21. Ibid. Assessment criteria undisclosed [requires primary verification]. ↩

  22. Forrester. “The State of Agentic Commerce in Mid-2026.” https://www.forrester.com/blogs/the-state-of-agentic-commerce-in-mid-2026/ ↩

  23. Forrester. “Predictions 2026: B2C Marketing, CX, & Digital Business.” https://www.forrester.com/press-newsroom/forrester-b2c-marketing-cx-digital-2026-predictions/ ↩

  24. McKinsey/QuantumBlack. “The State of AI.” Survey conducted June-July 2025, 1,993 respondents, 105 countries. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai ↩

  25. Deloitte. “State of AI in the Enterprise 2026.” 3,235 respondents, August-September 2025. https://www.deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html ↩

  26. Forrester Predictions 2026. Sample and period unclear from search snippet [requires primary verification]. ↩

  27. McGuinness, M. et al. 2026. “How we contain Claude across products.” https://www.anthropic.com/engineering/how-we-contain-claude ↩ ↩2

  28. Hadfield, J. et al. “How we built our multi-agent research system.” https://www.anthropic.com/engineering/multi-agent-research-system ↩

  29. Rajasekaran, P. “Harness design for long-running application development.” https://www.anthropic.com/engineering/harness-design-long-running-apps ↩ ↩2

  30. Chase, H. Sequoia Capital Podcast. ↩

  31. Lowry, T. “Why You Should Care About Design Context.” https://www.figma.com/blog/why-you-should-care-about-design-context/ ↩

  32. Rajasekaran, P. et al. “Effective context engineering for AI agents.” https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents ↩

  33. Field, D. Config 2026 keynote. https://www.youtube.com/watch?v=2ZCc4k_IV5w ↩

  34. Karpathy, A. 2026. Sequoia AI Ascent 2026. https://karpathy.bearblog.dev/sequoia-ascent-2026/ ↩

  35. Willison, S. 2025. “The lethal trifecta for AI agents.” https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/ ↩

  36. Willison, S. 2026. “Agentic Engineering Patterns.” https://simonwillison.net/2026/Feb/23/agentic-engineering-patterns/ ↩

← All Notes · Home