How Many Times Have AI Agents Been in Fashion over 70 Years, and What Did Design Do Each Time? (Academic Research)

The question and the scope of this note

Since around 2024, “AI agent” has become one of the most frequent terms in product announcements and in the titles of research papers. However, the term is not new. In the mid-1990s, agents were also described as the future of computing, and within a few years they were no longer discussed. If the current fashion is the third or fourth wave, it is worth knowing what happened in the earlier waves and why they receded.

This note asks two questions. The first is how the idea of the AI agent emerged, became fashionable, receded, and returned between the 1950s and 2026. The second is whether design (HCI, interaction design, and UI design) played a role in these rises and falls. A role here can mean creating the concept, giving it form, causing it to fail, rebuilding it, or arriving later to evaluate it. We decided to choose between the answer that design was always the protagonist and the answer that it was always a bystander only after checking the sources.

The material consists of 47 works: peer-reviewed papers and book chapters in AI and HCI, technical reports from academic institutions, and arXiv papers with a stated venue. arXiv papers without a peer-reviewed version are marked as such in the text. For papers whose full text could not be obtained (because the ACM or AAAI sites refused retrieval), only what the abstract states is used, and the text says “the abstract states”. Industry material (product announcements and vendor white papers) is not used. Recent research on human-agent interaction is covered in more detail in The Academic Foundations of Agentic Experience (AX): Where Human-Agent Interaction Research Stands.

Two sources: machines that think, and machines that work with people (1950s to 1970s)

Before the word agent spread, two different visions existed side by side. The first was a vision of making the machine itself think. In 1950, Turing replaced the question “Can machines think?” with an imitation game and devoted the last section of the paper to learning machines (Turing 1950). In 1959, McCarthy proposed the Advice Taker, a program that solves problems by manipulating sentences in formal languages (McCarthy 1959). The advantage McCarthy expected from this program was that its behavior could be improved merely by making statements to it, without knowledge of the program itself. The image of a machine that becomes smarter when told appears again in the later discussion of agents. In 1973, Hewitt and colleagues presented a formalism that builds computation from a single kind of object, the “actor”, which exchanges messages (Hewitt et al. 1973).

The second was a vision of people and computers thinking together. In 1960, Licklider wrote about “man-computer symbiosis” and distinguished it both from the mechanical extension of the human arm and eye and from artificial intelligence (Licklider 1960). Licklider avoided argument with enthusiasts for artificial intelligence by conceding that thinking in the distant future would belong to machines alone. He then wrote that during “a fairly long interim”, the main intellectual advances would be made by people and computers working together in intimate association. He estimated this period at 15 years and added that it might be 10 or 500.

Between these two visions, one problem was visible from the beginning. In 1966, Weizenbaum reported the conversational program ELIZA and wrote that once its workings are explained, its magic crumbles away (Weizenbaum 1966). At the same time, he wrote that some subjects had been very hard to convince that ELIZA was not human, and he stated that the credibility of machine output to humans itself demanded investigation. This gap between the simplicity of the mechanism and the impression people receive recurs in every later wave.

The agent became a word of the interface (1990s)

In 1983, Shneiderman formulated direct manipulation (Shneiderman 1983). His account consisted of a continuous representation of the object of interest, physical actions such as mouse movements and button presses in place of commands, and rapid, incremental, reversible operations whose results are immediately visible. The discussion of agents in the 1990s began as an alternative to this direct manipulation.

In 1994, Maes reported learning interface agents (Maes 1994). An interface agent is a program that uses artificial intelligence techniques to provide active assistance with a user’s computer-based tasks, and Maes explained it through the metaphor of a personal assistant. In direct manipulation, the user must initiate all tasks explicitly and monitor all events. Maes positioned the agent as implementing another style of interaction, which Kay had called “indirect management”. She then narrowed the problems to two. One is “competence”, namely how the agent knows when, with what, and how to help the user; the other is “trust”, namely whether the user feels comfortable delegating tasks to the agent. Maes’s answer was that by observing the user and learning gradually, the agent also gives the user time to understand gradually how the agent makes decisions.

In the same period, AI researchers were trying to define the agent. In 1995, Wooldridge and Jennings distinguished a “weak” notion, a computer system with autonomy, social ability, reactivity, and pro-activeness, from a “strong” notion described in terms of mental states such as beliefs and intentions (Wooldridge and Jennings 1995). In 1996, Franklin and Graesser defined an autonomous agent as a system situated within an environment that senses that environment and acts on it, over time, in pursuit of its own agenda, so as to affect what it senses in the future (Franklin and Graesser 1996). Earlier, in 1991, Brooks had presented an approach that builds intelligence by connecting to the environment through perception and action, without relying on representation (Brooks 1991).

The design side gave this word its form. Erickson, then at the User Experience Architects’ Office of Apple, pointed out that the word agent mixed two meanings (Erickson 1997). One is adaptive functionality, which works without direct control by a person, and the other is the agent metaphor, which portrays a program as a living being. As an example of the latter, Erickson cited the bow-tied human figure in Apple’s 1987 video Knowledge Navigator. In Apple’s Guides interface for an encyclopedia, guides with roles such as a settler woman and an inventor recommended related articles, but the guides were not given realistic human figures. Erickson concluded that research on how adaptive functionality is portrayed, rather than on the functionality itself, was lacking. In the 1994 CACM special issue in which Maes’s paper appeared, Bates drew on the work of Disney animators and discussed the expression of emotion as a requirement for “believable agents” (Bates 1994).

In the same period, the strongest criticism also came from the design side

The people who doubted agents most strongly were also people in design and HCI. Norman wrote that agents were surrounded by fear and extravagant claims and that the main difficulties he foresaw were social, not technical (Norman 1997). The issues Norman listed were the user’s feeling of control, safeguards against runaway computation, accurate expectations, privacy, and hiding complexity while revealing the underlying operations. He also wrote that users should be able to trace back what was done on their behalf and undo actions they did not want. He warned that giving an agent a human face or voice creates expectations of human-like intelligence and understanding, and thus overpromises.

In 1995, Lanier wrote that the idea of intelligent agents is both wrong and evil (Lanier 1995). Lanier’s argument was that people give extra deference to a program regarded as autonomous, fit their behavior to the agent’s categories in order to make the agent look smart, and as a result make themselves dumb. Lanier cited Microsoft Bob and Apple’s Newton as rare examples of shipped products that claimed agent capability.

Voices worried about the inflation of the word also came from the AI side. In the conclusion of his 1996 overview, Nwana wrote that “agent” had become a buzzword because it is both a technical concept and a metaphor (Nwana 1996). He then described the risk of repeating the path that artificial intelligence itself had followed, in which exaggerated claims drown out real achievements, interest wanes, and research funding disappears, and he recommended dropping the qualifier “intelligent”.

In 1997, Shneiderman and Maes held public debates at two conferences, IUI and CHI (Shneiderman and Maes 1997). Shneiderman set the goal of environments in which users comprehend the display, feel in control, find the system predictable, and are willing to take responsibility for their actions, and he stated that responsibility would be the central issue of the debate. He cited chatty bank tellers and the proposed Knowledge Navigator as examples of anthropomorphic interfaces, called Microsoft Bob a failed product, and stated that the anthropomorphic or social interface was not the future of computing. Maes replied that as environments become more complex, people will need to delegate to “extra eyes or extra ears”. She then argued that agents are not an alternative to direct manipulation but complementary metaphors, and that most agents are not personified. A year earlier, Gentner and Nielsen had also listed, as “Anti-Mac” principles that invert the Mac principles, delegation in place of direct manipulation and shared control in place of user control (Gentner and Nielsen 1996).

The debate was divided on more than whether to delegate. It also contested whether to portray agents in human form, who bears responsibility, and whether users can see and undo what was done. Maes’s rebuttal that agents need not be personified draws the same line as Erickson’s distinction between functionality and metaphor.

What happened with the Office assistant

In parallel with this controversy, agents entered a large product. The Office 97 assistant (commonly called Clippy) was based on Lumière, a research project at Microsoft Research (Horvitz et al. 1998). Lumière used Bayesian models to infer, from the stream of a user’s actions, what the user wanted to do and whether the user needed help.

The Lumière team worked with psychologists in Microsoft’s usability laboratory and conducted a “Wizard of Oz” study in which experts in a separate room watched only the user’s screen and typed advice. The experts identified users’ goals to some degree through observation, but recognizing goals was difficult, and they were uncertain about the value of different kinds of assistance. Another finding was that poor advice could be costly to users. Even though subjects had been told that the help system was experimental, they examined the advice carefully, became distracted by advice that was off the mark, and began to experiment with the features it described. Seeing this, the experts took it as false confirmation that their goal recognition had succeeded and pushed the users further down a distracting path.

The paper also describes what changed when the research became a product. The Office team committed to a character-based front end to relay the results of inference. The product’s models were broader but shallower than the research prototype, kept no persistent profile of user competence, and considered only a small number of recent actions. In addition, the mechanism that would provide assistance autonomously on the basis of the likelihood that a user needs help, or of the expected utility of such action, was not employed.

The following sequence can be read from this account. The usability study had found the cost of poor advice before the product. The research prototype offered help only when a probability threshold was exceeded, and the team had also experimented with controlling the threshold through cost-benefit analysis. Even so, the product shipped with a character in front and without the mechanism for deciding when to offer help. This can be read as a case in which knowledge from design and human factors existed but was not followed at the point of product decision. However, no peer-reviewed study that examined the failure of this product itself was found in this survey.

In the following year, 1999, Horvitz, the first author of Lumière, set out principles for mixed-initiative user interfaces and listed the Lumière paper among his references (Horvitz 1999). In response to the debate over direct manipulation versus agents, Horvitz sought principles for combining the two. The main problems of agents he listed were poor guessing about users’ goals, inadequate consideration of the costs and benefits of automated action, poor timing, and inadequate opportunities for users to guide automated services and refine their results. He also wrote that complex reasoning machinery should not be built to patch fundamentally poor designs and metaphors.

What remained while the word receded (2000s and 2010s)

In the 2000s, the word agent became less visible in the discussion of interfaces. However, the word had moved to other areas of computing and to fields outside computing. Niazi and Hussain conducted a scientometric analysis of 1,064 journal articles on agent-based computing indexed in the Web of Science from 1990 to 2010 (Niazi and Hussain 2011). Their results showed a strong presence of multi-agent systems and agent-based models in non-computing domains such as the life sciences, ecological sciences, and social sciences. In reinforcement learning, the agent also continued to denote a learner that tries to control an environment. The 2015 DQN paper reported a system that took only pixels and the game score as inputs and reached a level comparable to a professional human games tester across 49 games (Mnih et al. 2015). The authors described it as the first artificial agent capable of learning to excel at a diverse array of challenging tasks.

On the interface side, it was tools that remained. In 2000, Parasuraman, Sheridan, and Wickens presented a model that divides automation into four classes of functions, information acquisition, information analysis, decision and action selection, and action implementation, and places a continuum of levels from fully manual to fully automatic within each (Parasuraman et al. 2000). The premise of this model is that automation does not merely supplant human activity but changes it and can impose new coordination demands. In 2004, Lee and See organized misuse and disuse of automation as a problem of calibration, namely how well trust matches the true capabilities of the automation (Lee and See 2004). Both are works of human factors research and do not center on the word agent, but they became vocabulary for deciding what to entrust and how far.

Grudin’s 2009 essay offers an account of this period (Grudin 2009). According to the abstract, AI and HCI dealt with the same subject, computing and intelligent behavior, but AI pursued expensive long-term visions while HCI focused on short-term improvement of widely used hardware, and the two differed in priorities, methods, and assessment. The abstract states that, as a consequence, the two fields competed for resources, with HCI flourishing in AI winters and moving more slowly when AI was in favor. From this perspective, it may not be a coincidence that tools on the design side grew during the period when the word agent receded.

In the 2010s, agents returned in the form of voice assistants and chatbots. Luger and Sellen interviewed 14 users of conversational agents and reported in 2016 that user expectations were dramatically out of step with the operation of the systems (Luger and Sellen 2016). Grudin and Jacques looked back on the wave of chatbots in 2019 (Grudin and Jacques 2019). In 2016, task-focused chatbots became a centerpiece of machine intelligence and hundreds of thousands were built, but creating successful non-trivial chatbots proved more difficult than anticipated. In January 2018, Facebook shut down M, the chatbot that had helped inspire the surge. The two authors wrote that some developers then began to design human-chatbot (“humbot”) teams in which bots answer only the queries for which they have high confidence and pass the rest to humans.

Designers also found machine learning a difficult material. In 2017, Dove and colleagues surveyed UX practitioners, and their abstract states that machine learning had not produced the wealth of design innovation that other technologies had, possibly because it is a new and difficult design material (Dove et al. 2017). In 2020, Yang and colleagues synthesized cases from their own design, research, and teaching and narrowed the reasons to two (Yang et al. 2020). These are uncertainty about what AI can do and how well, and the complexity of AI outputs. Paper prototypes and Wizard of Oz prototypes had been tools that let HCI handle complex technologies cheaply. However, a system such as Siri, whose relationship between inputs and outputs cannot be described concisely, cannot be reproduced by rule-based simulators or by a human wizard, which makes it difficult to examine through prototyping how inference errors affect the experience. Yang and colleagues acknowledged as a limitation that their cases came mainly from their own experience and were not representative.

Guidelines were nevertheless compiled. In 2019, Amershi and colleagues consolidated more than 150 AI design recommendations from academic and industry sources into 18 guidelines, which 49 design practitioners tested against 20 AI products (Amershi et al. 2019). This paper cites Horvitz’s 1999 paper as the formative paper on mixed-initiative systems that balanced autonomous action with direct manipulation. In 2020, Shneiderman presented a framework that treats the level of automation and the level of human control as two separate axes rather than as the two ends of a single axis, with the goal of making both high (Shneiderman 2020). This can be read as a proposal to resolve the opposition of the 1997 debate by using two axes.

The return of LLM agents (2022 to 2026)

The return of agents after 2022 began on the side of machine learning and natural language processing. ReAct is a method that has a language model alternate between reasoning traces and actions, and on two interactive decision-making benchmarks (ALFWorld and WebShop) it outperformed imitation and reinforcement learning methods by an absolute success rate of 34% and 10% (Yao et al. 2023). Toolformer is a method in which a language model learns, in a self-supervised way, which APIs to call, when to call them, and what arguments to pass (Schick et al. 2023). Surveys by Wang and colleagues and by Xi and colleagues organized the components and applications of agents based on LLMs (Wang et al. 2024; Xi et al. 2025).

This return is not cut off from the past. The survey by Wang and colleagues quotes Franklin and Graesser’s 1996 definition of an autonomous agent verbatim at its opening. The survey by Xi and colleagues traces the concept of the agent from its philosophical origins to its development in AI before explaining why LLMs are suitable foundations for agents. Sumers and colleagues drew on the long history of cognitive science and symbolic artificial intelligence to propose a framework that describes language agents in terms of memory, action space, and decision-making procedure (Sumers et al. 2024). As the word spread rapidly, efforts to sort out its uses also began. Sapkota and colleagues distinguished a single “AI agent” that uses tools to automate well-defined tasks from “agentic AI”, in which multiple specialized agents divide and coordinate work (Sapkota et al. 2026).

The level of capability was measured in evaluation environments. In WebArena, the best GPT-4-based agent completed 14.41% of the tasks end to end, while humans completed 78.24% (Zhou et al. 2024). In OSWorld, which collects tasks on real operating systems, humans accomplished 72.36%, while the best model reached only 12.24%, struggling mainly with GUI grounding and operational knowledge (Xie et al. 2024). Criticism of evaluation itself also appeared. Kapoor and colleagues argued that agent evaluation focuses narrowly on accuracy without attention to cost, and that as a result needlessly complex and costly agents are being built (Kapoor et al. 2024, arXiv). Chan and colleagues organized four characteristics, underspecification, directness of impact, goal-directedness, and long-term planning, that increase the “agency” of algorithmic systems, together with harms that become systemic and long-range (Chan et al. 2023). Mitchell and colleagues analyzed that risks to people increase as users cede more control to an agent and argued that fully autonomous agents should not be developed (Mitchell et al. 2025, arXiv).

In this wave as well, design arrived after capability. There is one exception. Generative agents by Park and colleagues, one of the representative works on LLM agents, was presented at UIST, an HCI conference (Park et al. 2023). This work built a small town of 25 agents in which users can intervene in natural language, cited Bates’s 1994 “believable agents”, and set believable proxies of human behavior as its goal.

Design research since then has restated the issues of the 1990s. Feng and colleagues treated the level of an agent’s autonomy as a design decision made separately from its capability and environment, and defined five levels by the roles a user takes: operator, collaborator, consultant, approver, and observer (Feng et al. 2025, arXiv). Like the levels of automation of Parasuraman and colleagues, this approach decides in stages how far to entrust. Bansal and colleagues identified 12 challenges in conveying information from the agent to the user and from the user to the agent (Bansal et al. 2024, arXiv).

Cocoa by Feng and colleagues is an interface in which people and agents alternately share planning and execution, and it was designed from a formative study with 9 researchers (Feng et al. 2026). In a lab study with 16 researchers, Cocoa afforded greater agent steerability while maintaining ease of use compared with a strong chat baseline. In a 7-day field deployment with 7 researchers, the explicit delegation of work between people and the agent was evaluated as valuable.

Shome and colleagues reviewed the marketed use cases of 102 commercial agents and then had 31 participants attempt representative tasks with Operator and Manus (Shome et al. 2026). The users were impressed by the agents but faced significant usability challenges, because the agents’ capabilities were misaligned with users’ mental models and the agents lacked the meta-cognitive abilities needed for collaboration. In their introduction, Shome and colleagues cited Horvitz’s mixed initiative and the 1997 debate between Shneiderman and Maes and wrote that the dream has resurfaced under the name of LLM agents. Zhu and colleagues argued that the limited adoption of agents stems not only from technical capability but also from a lack of design knowledge for human-agent interaction, and applied 14 design principles to 9 systems (Zhu et al. 2026, arXiv).

Did design play a role?

The sources allow the answer that design did play a role. However, the role differed by period, and in no period was it the role of producing capability.

In the 1960s, the vision of machines that work with people (Licklider) became the counterpart of the vision of autonomous machines. This pairing remains, in changed forms, in the 1997 debate, in Shneiderman’s two axes of 2020, and in the levels of autonomy of Feng and colleagues in 2025. In the 1990s, the design side made the agent a word of the interface (Maes, Erickson, Bates), and the same side produced the strongest criticism (Norman, Lanier, Shneiderman). In other words, those who created the fashion and those who tried to stop it were in the same field. At the end of the 1990s, a usability study had found the risk before the product, but the product shipped in a form different from that knowledge (Horvitz et al. 1998). In the 2000s and 2010s, when the word receded, vocabulary for deciding what to entrust, how far, and how to present it grew on the side of design and human factors: levels of automation, calibration of trust, mixed initiative, and 18 guidelines. After 2022, capability came first from the machine learning side, and design has followed with evaluation and interface design. Its issues (the scope of delegation, portrayal, undo, and the mismatch with user expectations) have not moved far from Maes’s competence and trust (1994) and Norman’s feeling of control and management of expectations (1997).

Two points follow from this account. The first is that the contribution of design has remained in each period as vocabulary for deciding how capability is presented and how far it is entrusted. The second is that the literature in this survey does not allow us to confirm how far that vocabulary was followed in product decisions, except in the single case of the Office assistant. Nor can the rise and fall of the fashion be attributed to design. What Grudin’s abstract offers is an account based on competition for resources between AI and HCI, and no study was found showing that criticism from design caused the decline of agents.

What research has not yet answered

No bibliometric study was found that counted the fashion of the word agent with the same method from the 1990s to the 2020s. For 1990 to 2010 there is the analysis by Niazi and Hussain, but there is no comparison between interface agents and the “agentic AI” of the 2020s. No peer-reviewed study that examined the failure of the Office assistant with primary sources was found. Within the scope of this survey, no study was found that measured how the issues raised by the criticism of the 1990s (the feeling of control, undo, and the overpromising of human portrayal) are reflected in the interfaces of current agent products. Research in the Japanese-language sphere (the Japanese Society for Artificial Intelligence, the Human Interface Society, and the HCI special interest group of the Information Processing Society of Japan) was not examined this time.

Related notes

• The Academic Foundations of Agentic Experience (AX): Where Human-Agent Interaction Research Stands: research on human-agent interaction, 2024 to 2026
• Industry Discourse on Agentic Experience (AX): What Vendors, Consultants, and Practitioners Are Saying: the industry version of the same topic
• What Should Creators Be Able to Take Over When Generative AI Does the Making? (Lessons from Aviation Automation Research): the ability of creators to take work back from AI, seen from aviation automation research
• End-User Development (EUC/EUD): Genealogy and Limits — A Map of 152 Sources: research on users creating programs themselves

Unverified items

No unverified statements remain in the main claims. Because no peer-reviewed study of the failure of the Office assistant itself was found, the text describes only what the Lumière paper states.

References

• Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., Teevan, J., Kikin-Gil, R., & Horvitz, E. (2019). Guidelines for human-AI interaction. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, 1–13. https://doi.org/10.1145/3290605.3300233
• Bansal, G., Wortman Vaughan, J., Amershi, S., Horvitz, E., Fourney, A., Mozannar, H., Dibia, V., & Weld, D. S. (2024). Challenges in human-agent communication. arXiv:2412.10380. https://arxiv.org/abs/2412.10380
• Bates, J. (1994). The role of emotion in believable agents. Communications of the ACM, 37(7), 122–125. https://doi.org/10.1145/176789.176803
• Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence, 47(1–3), 139–159. https://doi.org/10.1016/0004-3702%2891%2990053-M
• Chan, A., Salganik, R., Markelius, A., Pang, C., Rajkumar, N., Krasheninnikov, D., Langosco, L., He, Z., Duan, Y., Carroll, M., et al. (2023). Harms from increasingly agentic algorithmic systems. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 651–666. https://doi.org/10.1145/3593013.3594033
• Dove, G., Halskov, K., Forlizzi, J., & Zimmerman, J. (2017). UX design innovation: Challenges for working with machine learning as a design material. Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, 278–288. https://doi.org/10.1145/3025453.3025739
• Erickson, T. (1997). Designing agents as if people mattered. In J. M. Bradshaw (Ed.), Software agents. AAAI Press / MIT Press. http://www.tomeri.org/agents.html
• Feng, K. J. K., McDonald, D. W., & Zhang, A. X. (2025). Levels of autonomy for AI agents. arXiv:2506.12469. https://arxiv.org/abs/2506.12469
• Feng, K. J. K., Pu, K., Latzke, M., August, T., Siangliulue, P., Bragg, J., Weld, D. S., Zhang, A. X., & Chang, J. C. (2026). Cocoa: Co-planning and co-execution with AI agents. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, 1–23. https://doi.org/10.1145/3772318.3791673
• Franklin, S., & Graesser, A. (1996). Is it an agent, or just a program?: A taxonomy for autonomous agents. In Intelligent Agents III (ATAL 1996), LNCS 1193, 21–35. Springer. https://doi.org/10.1007/BFb0013570
• Gentner, D., & Nielsen, J. (1996). The Anti-Mac interface. Communications of the ACM, 39(8), 70–82. https://doi.org/10.1145/232014.232032
• Grudin, J. (2009). AI and HCI: Two fields divided by a common focus. AI Magazine, 30(4), 48–57. https://doi.org/10.1609/aimag.v30i4.2271
• Grudin, J., & Jacques, R. (2019). Chatbots, humbots, and the quest for artificial general intelligence. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, 1–11. https://doi.org/10.1145/3290605.3300439
• Hewitt, C., Bishop, P., & Steiger, R. (1973). A universal modular ACTOR formalism for artificial intelligence. Proceedings of the 3rd International Joint Conference on Artificial Intelligence, 235–245. https://www.ijcai.org/Proceedings/73/Papers/027B.pdf
• Horvitz, E. (1999). Principles of mixed-initiative user interfaces. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ‘99), 159–166. https://doi.org/10.1145/302979.303030
• Horvitz, E., Breese, J., Heckerman, D., Hovel, D., & Rommelse, K. (1998). The Lumière project: Bayesian user modeling for inferring the goals and needs of software users. Proceedings of the 14th Conference on Uncertainty in Artificial Intelligence (UAI 1998), 256–265. https://arxiv.org/abs/1301.7385
• Kapoor, S., Stroebl, B., Siegel, Z. S., Nadgir, N., & Narayanan, A. (2024). AI agents that matter. arXiv:2407.01502. https://arxiv.org/abs/2407.01502
• Lanier, J. (1995). Agents of alienation. interactions, 2(3), 66–72. https://doi.org/10.1145/208666.208684
• Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50.30392
• Licklider, J. C. R. (1960). Man-computer symbiosis. IRE Transactions on Human Factors in Electronics, HFE-1(1), 4–11. https://doi.org/10.1109/THFE2.1960.4503259
• Luger, E., & Sellen, A. (2016). “Like having a really bad PA”: The gulf between user expectation and experience of conversational agents. Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, 5286–5297. https://doi.org/10.1145/2858036.2858288
• Maes, P. (1994). Agents that reduce work and information overload. Communications of the ACM, 37(7), 30–40. https://doi.org/10.1145/176789.176792
• McCarthy, J. (1959). Programs with common sense. In Mechanisation of Thought Processes: Proceedings of the Symposium of the National Physical Laboratory. HMSO. http://www-formal.stanford.edu/jmc/mcc59.pdf
• Mitchell, M., Ghosh, A., Luccioni, A. S., & Pistilli, G. (2025). Fully autonomous AI agents should not be developed. arXiv:2502.02649. https://arxiv.org/abs/2502.02649
• Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M., Fidjeland, A. K., Ostrovski, G., et al. (2015). Human-level control through deep reinforcement learning. Nature, 518(7540), 529–533. https://doi.org/10.1038/nature14236
• Niazi, M., & Hussain, A. (2011). Agent-based computing from multi-agent systems to agent-based models: A visual survey. Scientometrics, 89(2), 479–499. https://doi.org/10.1007/s11192-011-0468-9
• Norman, D. A. (1997). How might people interact with agents. In J. M. Bradshaw (Ed.), Software agents. AAAI Press / MIT Press. (Revised from Communications of the ACM, 37(7), 68–71, 1994.) https://web.archive.org/web/2005/http://www.jnd.org/dn.mss/agents.html
• Nwana, H. S. (1996). Software agents: An overview. The Knowledge Engineering Review, 11(3), 205–244. https://doi.org/10.1017/S026988890000789X
• Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 30(3), 286–297. https://doi.org/10.1109/3468.844354
• Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST ‘23), 1–22. https://doi.org/10.1145/3586183.3606763
• Sapkota, R., Roumeliotis, K. I., & Karkee, M. (2026). AI agents vs. agentic AI: A conceptual taxonomy, applications and challenges. Information Fusion, 126, 103599. https://doi.org/10.1016/j.inffus.2025.103599
• Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., & Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems 36 (NeurIPS 2023). https://arxiv.org/abs/2302.04761
• Shneiderman, B. (1983). Direct manipulation: A step beyond programming languages. Computer, 16(8), 57–69. https://doi.org/10.1109/MC.1983.1654471
• Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. International Journal of Human–Computer Interaction, 36(6), 495–504. https://doi.org/10.1080/10447318.2020.1741118
• Shneiderman, B., & Maes, P. (1997). Direct manipulation vs. interface agents. interactions, 4(6), 42–61. https://doi.org/10.1145/267505.267514
• Shome, P., Krishnan, S., & Das, S. (2026). Why Johnny can’t use agents: Industry aspirations vs. user realities with AI agents. Proceedings of the ACM Conference on AI and Agentic Systems (CAIS ‘26), 597–619. https://doi.org/10.1145/3786335.3813140
• Sumers, T. R., Yao, S., Narasimhan, K., & Griffiths, T. L. (2024). Cognitive architectures for language agents. Transactions on Machine Learning Research. https://arxiv.org/abs/2309.02427
• Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460. https://doi.org/10.1093/mind/LIX.236.433
• Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., et al. (2024). A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6). https://doi.org/10.1007/s11704-024-40231-1
• Weizenbaum, J. (1966). ELIZA: A computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36–45. https://doi.org/10.1145/365153.365168
• Wooldridge, M., & Jennings, N. R. (1995). Intelligent agents: Theory and practice. The Knowledge Engineering Review, 10(2), 115–152. https://doi.org/10.1017/S0269888900008122
• Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al. (2025). The rise and potential of large language model based agents: A survey. Science China Information Sciences, 68(2). https://doi.org/10.1007/s11432-024-4222-0
• Xie, T., Zhang, D., Chen, J., Li, X., Zhao, S., Cao, R., Hua, T. J., Cheng, Z., Shin, D., Lei, F., et al. (2024). OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments. Advances in Neural Information Processing Systems 37 (NeurIPS 2024), Datasets and Benchmarks Track. https://arxiv.org/abs/2404.07972
• Yang, Q., Steinfeld, A., Rosé, C., & Zimmerman, J. (2020). Re-examining whether, why, and how human-AI interaction is uniquely difficult to design. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 1–13. https://doi.org/10.1145/3313831.3376301
• Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2023). ReAct: Synergizing reasoning and acting in language models. International Conference on Learning Representations (ICLR 2023). https://arxiv.org/abs/2210.03629
• Zhou, S., Xu, F. F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Ou, T., Bisk, Y., Fried, D., Alon, U., & Neubig, G. (2024). WebArena: A realistic web environment for building autonomous agents. International Conference on Learning Representations (ICLR 2024). https://arxiv.org/abs/2307.13854
• Zhu, H., Wang, C., Xiao, Q., & Shen, H. (2026). Design principles for human-agent interaction. arXiv:2606.20630. https://arxiv.org/abs/2606.20630