Why Johnny Can't Use Agents: Industry Aspirations vs. User Realities with AI Agents

This paper investigates the gap between industry marketing and user realities of AI agents by analyzing 102 commercial tools and conducting a usability study with 31 participants, revealing that while users are impressed, they face significant challenges due to misaligned capabilities and a lack of meta-cognitive collaboration skills.

The Gist

Imagine you've just bought a brand-new, high-tech robot butler. The company's commercials show it doing everything perfectly: planning your entire vacation, building a slide deck for your boss, and researching your next career move, all while you sip coffee and relax. The robot is marketed as an "AI Agent"—a smart partner that takes initiative and gets things done for you.

But when you actually turn it on and try to use it, things get messy. You might find yourself confused, frustrated, or unsure if the robot is actually helping or just making a bigger mess.

This paper, titled "Why Johnny Can't Use Agents," investigates exactly that gap between the shiny marketing promises of AI agents and the confusing reality of using them today. The researchers asked two main questions:

1. What are companies actually selling? (The Hype)
2. What happens when regular people try to use them? (The Reality)

Here is a breakdown of their findings using simple analogies.

1. The Three Types of "Robot Butlers" (The Hype)

The researchers looked at 102 different products sold as "AI Agents" and sorted them into three buckets based on what the companies say they do:

• The Orchestrator (The Travel Agent): These agents are supposed to go out, click buttons on websites, book flights, and fill out forms for you. They "orchestrate" a series of actions in the real world.
• The Creator (The Artist): These agents are supposed to make things for you, like slide decks, websites, or documents. They focus on the final product's look and format.
• The Insight Generator (The Researcher): These agents are supposed to dig through the internet, find information, and give you a summary or a recommendation. They are your personal librarian and analyst.

2. The Experiment: Putting "Johnny" to the Test

To see if these robots actually work, the researchers recruited 31 regular people (they call this persona "Johnny," a nod to an old study about why regular people couldn't use encryption). These participants were familiar with chatbots but had never used an AI agent that could control a computer.

They gave "Johnny" three specific tasks:

• Orchestration: Plan a 3-day holiday trip (booking flights and hotels).
• Creation: Make a 10-minute presentation slide deck.
• Insight: Figure out how to spend a $2,000 budget for personal growth.

They used two popular commercial agents (named Operator and Manus) to see how the humans fared.

3. The Five Big Problems (The Reality)

Even though the participants were generally impressed by the technology and could often finish the tasks, they hit five major walls that made the experience frustrating.

Barrier 1: The "Mind-Reading" Misunderstanding

The Analogy: Imagine you hire a new assistant. You say, "Make me a sandwich." You expect a ham sandwich. The assistant brings you a bowl of flour and a knife because they didn't know you wanted ham. You get annoyed, but you realize you didn't specify "ham."
The Reality: Users didn't know how much detail to give the AI. Some thought they had to write a perfect, step-by-step manual for the robot. Others thought the robot could read their mind. Because the AI didn't explain how it was thinking, users felt like they were "gambling" with their first prompt. If they got it wrong, the robot would go down the wrong path, and the user felt trapped.

Barrier 2: The "Trust Me" Leap

The Analogy: You ask a stranger to hold your wallet while you tie your shoe. They say, "I'll be right back," and run off with your wallet. You feel unsafe.
The Reality: The AI agents often asked for sensitive things (like logging into your Google account) or started making decisions (like booking a hotel) without asking, "Do you want a room with a pool or a view?" Users felt they had to trust the robot blindly, but the robot didn't earn that trust by explaining its choices or asking for permission first.

Barrier 3: The "One-Size-Fits-All" Dance Partner

The Analogy: Imagine dancing with a partner who only knows one style of dance. If you want to waltz, they try to breakdance. If you want to stop, they keep spinning.
The Reality: People have different styles of working. Some want to do the heavy lifting and just check the AI's work; others want the AI to do everything. The agents were too eager to just "do the job" without checking in. If a user wanted to pause or change the plan, the agent often didn't listen or made it hard to stop, leaving the user feeling like they had lost control of the dance.

Barrier 4: The "Firehose" of Information

The Analogy: You ask a friend for directions. Instead of saying "Turn left," they give you a 20-minute lecture on the history of the street, the traffic patterns, and the weather, while you're trying to drive.
The Reality: The agents were very chatty. They showed every single step they took, every search result, and every thought process. For some users, this was helpful; for others, it was overwhelming noise. It was hard to find the important parts because the "logs" were too dense and confusing.

Barrier 5: The Robot That Doesn't Know It's Stuck

The Analogy: You ask a GPS to find a route. It gets stuck in a loop, trying to drive through a wall, and keeps saying "Recalculating" without ever telling you, "Hey, I can't get through here, you need to drive manually."
The Reality: When the AI got stuck (like trying to log into a website that blocked robots), it often didn't realize it was failing. It would just freeze or repeat the same action over and over. It lacked the "self-awareness" to say, "I'm stuck, please help me." Users had to figure out the error themselves, which defeated the purpose of having an agent.

The Bottom Line

The paper concludes that while AI agents are powerful and can do amazing things, they aren't ready for prime time with regular people yet.

The technology is like a race car engine that hasn't been put into a car with a steering wheel, brakes, or a dashboard. The industry is selling the engine (the ability to do tasks), but users need the car (the ability to control, trust, and understand the engine).

Until these agents can better understand human expectations, explain their mistakes, and let us take the wheel when things go wrong, "Johnny" will keep struggling to use them effectively.

Technical Summary

Technical Summary: Why Johnny Can't Use Agents: Industry Aspirations vs. User Realities with AI Agents

Problem Statement
The paper addresses a growing imprecision regarding the definition, capabilities, and usability of "AI agents." While the technology industry markets these systems as intelligent partners capable of autonomous, multi-step execution, there is a lack of systematic understanding regarding how end-users actually interact with them. Prior evaluations of AI agents have largely focused on technical benchmarks and quantifiable ideals (e.g., task completion rates in controlled environments), often overlooking the human factors of delegation, oversight, and recovery. The authors posit that marketed capabilities often diverge from user realities, creating friction that prevents effective adoption by novice users. The core problem is the gap between industry aspirations (what agents are marketed to do) and user realities (the challenges faced when attempting to use them for advertised tasks).

Methodology
The research employs a two-pronged approach to investigate the disconnect between industry framing and user experience:

1. Systematic Review (RQ1): The authors constructed a taxonomy of marketed AI agent capabilities by analyzing $N=102$ commercial products sourced from aggregator directories (e.g., AI Agent Directory, Product Hunt) and web searches. They performed an inductive qualitative content analysis on marketing materials to distill advertised use cases into three broad categories: Orchestration (acting in GUIs on the user's behalf), Creation (generating structured artifacts like slides or code), and Insight (supporting research, synthesis, and recommendations).

2. Usability Assessment (RQ2): The authors conducted a think-aloud usability study with $N=31$ participants. Participants were novices to operationally agentic systems but frequent users of generative AI chatbots. They attempted representative tasks from each of the three taxonomy categories using two popular commercial operationally agentic platforms: OpenAI Operator and Manus.

   • Tasks: Holiday Planning (Orchestration), Slide Making (Creation), and Professional/Personal Growth Stipend Budgeting (Insight).
   • Procedure: Each session lasted approximately one hour, consisting of two 20-minute task attempts followed by semi-structured interviews. The study collected screen/audio recordings, System Usability Scale (SUS) scores, and interview transcripts.
   • Analysis: Data was analyzed using reflexive thematic analysis to identify recurring barriers and usability challenges.

Key Contributions
The paper makes three primary contributions to the field of Human-Computer Interaction (HCI) and AI:

1. A Taxonomy of Marketed Capabilities: A distilled framework categorizing industry-envisioned AI agent use cases into Orchestration, Creation, and Insight, clarifying how the "agent" label is currently applied in the commercial market.
2. Empirical Identification of Usability Barriers: An account of five critical usability barriers that novice users face when interacting with commercial AI agents, moving beyond simple task completion metrics to evaluate the quality of the delegation and collaboration process.
3. Design and Evaluation Implications: A set of concrete implications for designing and evaluating agentic systems, including specific axes for assessment (e.g., intervention frequency, time-to-recovery, stall/loop rate) that complement existing technical benchmarks.

Key Results and Findings
While participants were generally successful in completing the assigned tasks and reported high System Usability Scale (SUS) scores (indicating general impression of utility), the study revealed significant friction points that hinder optimal use. The authors identified five critical usability barriers:

1. Mental-Model Misalignment: Users struggled to understand the agent's capabilities, the required level of detail in prompts, and the agent's role during execution. This led to "prompt gambling" (uncertainty about how much to specify) and confusion regarding interaction mechanics like "Take Over" (user intervention). Users built mental models reactively from outcomes rather than proactively from system cues.
2. Premature Trust Assumptions: Agents often presumed trust in sensitive contexts (e.g., handling credentials, making travel plans) without establishing credibility or confirming user intent. Users expressed distrust regarding hallucinations, password management, and the agent's tendency to act without clarifying personal preferences.
3. Collaboration-Style Mismatch: Agents failed to accommodate diverse collaboration styles. Some users desired deep involvement and fine-grained control (acting as "thought partners"), while others wanted minimal involvement. Agents tended to be over-eager execution tools, assuming users wanted minimal oversight, and lacked mechanisms for effective mid-task steering or recovery from errors.
4. Communication Overload: Users faced difficulties parsing agent outputs. There was a spectrum of preferences regarding progress visibility; some found detailed logs overwhelming, while others felt they lacked necessary oversight. The communication overhead often made it difficult to articulate intent or identify where the agent was in the workflow.
5. Weak Metacognitive Behavior: Agents lacked the ability to self-assess their progress, limitations, or output quality. When agents encountered errors or stalls, they often failed to recognize the blockage, leading to repetitive loops or silent failures. Users were forced to cover these meta-cognitive gaps, often struggling to recover from opaque failure modes.

Significance and Claims
The paper claims that the transition from chat-based interaction to operationally agentic systems fundamentally changes the usability surface. In chatbots, a poor prompt may result in a suboptimal text response; in agents, the same ambiguity can trigger time-consuming, resource-intensive multi-step executions with real-world side effects (e.g., booking flights, modifying files) before the user can intervene.

The authors argue that structural requirements for agentic systems—delegation, oversight, intervention, and recovery—cannot be solved merely by expecting more capable users or more powerful models. Instead, the design of these systems must explicitly address the identified barriers by:

• Calibrating to user preferences regarding proactivity and communication.
• Improving agent self-assessment and transparency (e.g., exposing confidence, detecting stalls).
• Supporting non-textual inputs and precise iteration mechanisms.
• Redefining evaluation metrics to include human-centric dimensions like intervention frequency and time-to-recovery.

The study concludes that while current agents show promise, significant usability gaps remain between industry aspirations and the realities of novice end-users, necessitating a shift in design focus from pure capability to collaborative reliability.