Shuichiro Ogawa
日本語

Notes · updated 2026-06-27

The Academic Foundations of Agentic Experience (AX): Where Human-Agent Interaction Research Stands

In the Design in Tech Report 2026, John Maeda proposed “the transition from UX to AX (Agentic Experience)” (Maeda 2026). The claim is that the design question shifts from “how will users behave” to “how do we judge whether the agent did well.”

Yet as of June 2026, peer-reviewed papers written under the term “Agentic Experience” itself barely exist. The academic community studies the same problems under adjacent concepts such as human-agent interaction, AI delegation, and agentic AI UX. This note organizes the findings that form the academic foundations of AX, drawing on 36 academic sources (including 22 peer-reviewed papers and 10 preprints)1.

Related: agentic-experience-industry (industry discourse) / agentic-experience-design-synthesis (academia x industry comparison) / design-agent-tools-landscape-2026 (industry tool trends) / vibe-coding-design-production (UI-generating agents) / mcp-design-agent-integration (MCP integration) / the roundtable on delivering multi-agent value (the value of multi-agent systems).

Overview of the Review Papers

In areas adjacent to AX, seven review papers have appeared since 2024.

Zhang et al. (2025) conducted a scoping review of 134 papers spanning 20 years from CHI, CSCW, and UIST, systematizing agency distribution patterns in human-AI co-creation2. As an organization of the patterns of who holds control at which stage, it provides the foundation for AX’s design problems.

Zhou, Corbett et al. (2026) compared human-agent and human-human interactions in a meta-analysis of 162 studies3. The finding that prosocial behavior and moral engagement decline in interactions with agents while task performance and trust remain equivalent suggests an asymmetry for AX design: “efficiency on par with humans, but degraded social and ethical quality.”

Zou et al. (2025, accepted to ACL 2026 Findings) is a survey of LLM-based human-agent collaboration systems, covering environment profiling, feedback mechanisms, interaction types, orchestration, and communication protocols across the board4.

Stige et al. (2024), an SLR of 46 papers, mapped AI use across the entire UX design process (context understanding, requirements elicitation, design assistance, evaluation, development) and identified a research gap around creativity tasks in human-AI collaboration5. Luo et al. (2025), an SLR of 83 papers, analyzed AI-enabled UX tools, reporting that AI is used most heavily to make UX evaluation more efficient, but also that junior designers risk stopping at developing prompting ability without honing design skills6.

Agency Distribution: Who Controls What, and When

The central design problem of AX is “the distribution of agency.” The design of how much autonomy to grant the agent and where humans intervene determines the quality of the experience.

Zhou, Roy et al. (2026, CHI 2026) derived the optimal checking frequency for multi-step agent tasks with a deterministic model7. The results were clear: 81% of participants preferred the intermediate-checking approach, and task completion rates improved by roughly 13.5%. The finding that checking at appropriate intervals is optimal — neither full autonomy (no checking) nor confirming every step — provides a basic parameter for AX design.

Adam et al. (2024) tested the difference between AI systems autonomously assigning tasks (technology-invoked) and users delegating on their own initiative (user-invoked) in two online experiments8. AI-initiated assignment increases users’ sense of “self-threat” and lowers acceptance of delegation. The finding that perceived control has a buffering effect demonstrates that the more autonomously an agent acts, the more design care is required.

Houtti et al. (2025, CHI 2025) proposed the OAI framework (Observe → Ask → Intervene) for AI agents in meetings9. In an evaluation with 68 users, participants preferred the “question-first” style (the agent asks first and acts only after obtaining the user’s permission) over fully autonomous intervention. This result shows that designing the “Ask” stage of agency distribution is the key to AX.

Goyal, Chang, and Terry (2024, CHI EA) identified six dimensions of human-agent alignment (knowledge schema, autonomy, operation, reputation, ethics, engagement)10. By making explicit the multidimensionality of the conditions an agent needs in order to execute tasks “properly,” it shows that AX is not a simple automation problem.

The Trade-off Between Trust and Oversight

The greater an agent’s autonomy, the more the design of trust and oversight becomes the problem.

He, Demartini, and Gadiraju (2025, CHI 2025) tested the “Plan-Then-Execute” paradigm in an experiment with 248 participants11. Presenting high-quality plans improves trust and performance, but the study also demonstrated an overtrust risk toward superficially plausible plans. There is a dilemma: an agent presenting its plan raises transparency, but users are then required to have the ability to judge that plan’s quality.

Yu et al. (2025, CSCW 2025) used an AI design agent (poster generation) to experiment with the effect of process transparency (disclosing the chain-of-thought) on UX12. In the high-transparency condition, trust, satisfaction, and willingness to use improved significantly. This result gives the “humans judge” design principle of the companies described in design-agent-tools-landscape-2026 a concrete measure: “make the process visible so that judgment is possible.”

Grunde-McLaughlin et al. (2026, preprint, Microsoft Research) identified the limits of current designs for agents’ reasoning and action traces through three user studies13. Improved interfaces shortened error-detection time, but gains in final accuracy were limited. The observation that improving oversight UIs does not eliminate the structural problem of users missing errors raises a fundamental challenge for AX design.

The Gap Between Industry Promotion and User Reality

Shome, Krishnan, and Das (2026, CAIS 2026) quantified the gap between industry promotion and actual UX through a systematic review of 102 commercial AI agents and usability testing with 31 users14. They identify the market categories of commercial agents (three types: orchestration, creation, insight) and the mental-model mismatches and lack of metacognitive capability that users actually experience.

This finding echoes Willison’s and Osmani’s “watershed between vibe coding and AI-assisted engineering” organized in vibe-coding-design-production. The gap between the tool’s capability (what it can do) and the user’s capability (what they can judge) is the essential design challenge of AX.

Naik et al. (2025, DIS 2025), through interviews with 13 multi-agent GenAI developers at Microsoft, extracted three challenges: managing complexity, promoting transparency, and balancing autonomy with human oversight15. The same group’s follow-up work (2026, preprint) redefines transparency as a multidimensional sociotechnical practice with five facets: reproducibility, debugging, boundary-setting, visualization, and auditing16.

Proposed Design Frameworks

Multiple frameworks have been proposed to systematize the design of AX.

Zhu et al. (2026, preprint, CMU) proposed 14 design principles spanning four interaction stages (initial engagement, continuation, sustained use, failure) and validated their utility on nine agent systems17. Their claim that the barrier to agent adoption lies not in technology but in the lack of design knowledge shows that AX is a problem of design, not of engineering.

D’Oro et al.’s (2025, preprint) ADEPTS framework is an attempt to consolidate scattered HAI design guidance (UX heuristics, engineering taxonomies, ethics checklists) into six principles18.

Borghoff et al. (2025) reconceptualized HCI as dynamic interaction among networked human and computational agents, proposing a formal framework using Petri nets19. Its distinction between multi-agent systems and centaur-type (human-AI integrated) systems gives theoretical grounding to the “distinction between convergent and non-convergent multi-agent systems” discussed in the roundtable on delivering multi-agent value.

Effects on Junior Designers

Multiple studies raise concerns about the effects of AI agents on junior designers.

In Li et al.’s (2024, CHI 2024) interviews with 20 designers, experienced designers viewed GenAI as an auxiliary tool and remained confident in their creativity and capacity for empathy, whereas junior designers worried about the risks of skill deterioration and creative depletion20. Luo et al.’s (2025) SLR likewise reports the structural risk that juniors stop at developing prompting ability without honing design skills.

This concern is consistent with Dylan Field’s observation in design-agent-tools-landscape-2026 that “AI has lowered the floor but not raised the ceiling.” A lower floor lowers the barrier to entry, but there is a risk that AI short-circuits the path for developing the abilities that raise the ceiling (design judgment, quality assessment, aesthetic judgment).

Unresolved Issues

Through this review, the following unresolved issues emerge in AX research.

  1. Optimal checking frequency is context-dependent: Zhou et al.’s (2026) “intermediate checking is optimal” is a result for specific tasks; in domains with heavy subjective judgment, such as design production, the optimal frequency may differ.

  2. The transparency paradox: High transparency increases trust (Yu et al. 2025), but improving oversight UIs yields only limited gains in accuracy (Grunde-McLaughlin et al. 2026). Raising transparency alone does not solve AX’s trust problem.

  3. Juniors’ learning pathways: There is a structure in which the more agents take over execution, the fewer opportunities juniors have to learn design judgment. Here lies the academic backing for the “the better the learning works, the more users churn” dilemma discussed in the roundtable on delivering multi-agent value.

  4. The academic standing of “AX”: Maeda’s proposal is a concept introduced by a practitioner, and the academic community advances the research within the “human-agent interaction” framework. Whether AX takes hold as an independent academic concept remains undecided. The CHI 2026 workshop “Agentic Automation Experiences” (DOI: 10.1145/3772363.3778732) is its germinal form.

References

  • Zhang, S. et al. 2025. “Exploring Collaboration Patterns and Strategies in Human-AI Co-creation through the Lens of Agency.” Proc. ACM HCI (CSCW). DOI: 10.1145/3757594
  • Zhou, J., Corbett, F. et al. 2026. “Psychological and behavioural responses in human-agent vs. human-human interactions.” Communications Psychology (Nature). DOI: 10.1038/s44271-026-00466-z
  • Zhou, J., Roy, A. et al. 2026. “When Should Users Check?” CHI 2026. DOI: 10.1145/3772318.3790655
  • Shome, P. et al. 2026. “Why Johnny Can’t Use Agents.” CAIS 2026. arXiv: 2509.14528
  • Zou, H. P. et al. 2025. “LLM-Based Human-Agent Collaboration and Interaction Systems: A Survey.” ACL 2026 Findings. arXiv: 2505.00753
  • He, G. et al. 2025. “Plan-Then-Execute.” CHI 2025. arXiv: 2502.01390
  • Houtti, M. et al. 2025. “Observe, Ask, Intervene.” CHI 2025. DOI: 10.1145/3706598.3713838
  • Epperson, W. et al. 2025. “Interactive Debugging and Steering of Multi-Agent AI Systems.” CHI 2025. DOI: 10.1145/3706598.3713581
  • Naik, S. et al. 2025. “Designing with Multi-Agent Generative AI.” DIS 2025. DOI: 10.1145/3715336.3735823
  • Goyal, N. et al. 2024. “Designing for Human-Agent Alignment.” CHI EA 2024. DOI: 10.1145/3613905.3650948
  • Adam, M. et al. 2024. “Navigating autonomy and control in human-AI delegation.” Decision Support Systems, 180. DOI: 10.1016/j.dss.2024.114193
  • Yu, A. et al. 2025. “Exploring the Impact of Process Transparency on UX in AI Design Agents.” CSCW Comp. 2025. DOI: 10.1145/3715070.3749256
  • Borghoff, U. M. et al. 2025. “Human-Artificial Interaction in the Age of Agentic AI.” Frontiers in Human Dynamics, 7. DOI: 10.3389/fhumd.2025.1579166
  • Stige, Å. et al. 2024. “AI for UX design: a systematic literature review.” Information Technology & People, 37(6). DOI: 10.1108/ITP-07-2022-0519
  • Luo, J. et al. 2025. “Designing With AI: A SLR on AI-Enabled UX Design Tools.” Advances in HCI. DOI: 10.1155/ahci/3869207
  • Li, J. et al. 2024. “UX Design Professionals’ Perceptions of Generative AI.” CHI 2024. DOI: 10.1145/3613904.3642114
  • Zhu, H. et al. 2026. “Design Principles for Human-Agent Interaction.” arXiv: 2606.20630 [requires primary verification: not peer-reviewed]
  • D’Oro, P. et al. 2025. “ADEPTS: A Capability Framework for Human-Centered Agent Design.” arXiv: 2507.15885 [requires primary verification: not peer-reviewed]
  • Grunde-McLaughlin, M. et al. 2026. “Overseeing Agents Without Constant Oversight.” arXiv: 2602.16844 [requires primary verification: not peer-reviewed]
  • Naik, S. et al. 2026. “Catch-22: Transparency in Multi-Agent LLM Systems.” arXiv: 2606.08323 [requires primary verification: not peer-reviewed]
  • 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 (non-peer-reviewed, practitioner)

Footnotes

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

  2. Zhang, S. et al. (2025). Exploring Collaboration Patterns and Strategies in Human-AI Co-creation through the Lens of Agency. Proc. ACM HCI (CSCW). DOI: 10.1145/3757594 ↩

  3. Zhou, J. et al. (2026). Psychological and behavioural responses in human-agent vs. human-human interactions. Communications Psychology (Nature). DOI: 10.1038/s44271-026-00466-z ↩

  4. Zou, H. P. et al. (2025). LLM-Based Human-Agent Collaboration and Interaction Systems: A Survey. ACL 2026 Findings. arXiv: 2505.00753 ↩

  5. Stige, Å. et al. (2024). AI for UX design: a systematic literature review. Information Technology & People, 37(6). DOI: 10.1108/ITP-07-2022-0519 ↩

  6. Luo, J. et al. (2025). Designing With AI: A SLR on AI-Enabled UX Design Tools. Advances in HCI. DOI: 10.1155/ahci/3869207 ↩

  7. Zhou, J., Roy, A. et al. (2026). When Should Users Check? CHI 2026. DOI: 10.1145/3772318.3790655 ↩

  8. Adam, M. et al. (2024). Navigating autonomy and control in human-AI delegation. Decision Support Systems, 180. DOI: 10.1016/j.dss.2024.114193 ↩

  9. Houtti, M. et al. (2025). Observe, Ask, Intervene. CHI 2025. DOI: 10.1145/3706598.3713838 ↩

  10. Goyal, N. et al. (2024). Designing for Human-Agent Alignment. CHI EA 2024. DOI: 10.1145/3613905.3650948 ↩

  11. He, G. et al. (2025). Plan-Then-Execute. CHI 2025. arXiv: 2502.01390 ↩

  12. Yu, A. et al. (2025). Exploring the Impact of Process Transparency on UX in AI Design Agents. CSCW Comp. 2025. DOI: 10.1145/3715070.3749256 ↩

  13. Grunde-McLaughlin, M. et al. (2026). Overseeing Agents Without Constant Oversight. arXiv: 2602.16844 [requires primary verification: not peer-reviewed] ↩

  14. Shome, P. et al. (2026). Why Johnny Can’t Use Agents. CAIS 2026. arXiv: 2509.14528 ↩

  15. Naik, S. et al. (2025). Designing with Multi-Agent Generative AI. DIS 2025. DOI: 10.1145/3715336.3735823 ↩

  16. Naik, S. et al. (2026). Catch-22: Transparency in Multi-Agent LLM Systems. arXiv: 2606.08323 [requires primary verification: not peer-reviewed] ↩

  17. Zhu, H. et al. (2026). Design Principles for Human-Agent Interaction. arXiv: 2606.20630 [requires primary verification: not peer-reviewed] ↩

  18. D’Oro, P. et al. (2025). ADEPTS: A Capability Framework for Human-Centered Agent Design. arXiv: 2507.15885 [requires primary verification: not peer-reviewed] ↩

  19. Borghoff, U. M. et al. (2025). Human-Artificial Interaction in the Age of Agentic AI. Frontiers in Human Dynamics, 7. DOI: 10.3389/fhumd.2025.1579166 ↩

  20. Li, J. et al. (2024). UX Design Professionals’ Perceptions of Generative AI. CHI 2024. DOI: 10.1145/3613904.3642114 ↩

← All Notes · Home