Why Johnny can’t use agents
Industry aspirations versus user realities with AI agent software.
An investigation into delegated work.
There is growing imprecision about what “AI agents” are, what they can do, and how effectively they can be used by their intended users. We pose two research questions: (i) how does the tech industry conceive of and market “AI agents,” and (ii) what challenges do end-users face when attempting to use commercial AI agents for their advertised uses?
We performed a systematic review of marketed use cases for 102 commercial AI agents, finding that they fall into three umbrella categories — orchestration, creation, and insight. We then observed N = 31 participants attempting representative tasks for each category on two popular commercial AI agent tools, surfacing five usability barriers and six design recommendations.
1.A mixed-methods approach.
Our study proceeds in two parts. First, a systematic review of commercial agents yields a taxonomy of marketed use cases. Second, a think-aloud and interview study observes real users attempting representative tasks drawn from that taxonomy.
2.What the industry sells & how well users can use it.
2.1Taxonomy of commercial AI agents
Across 102 reviewed tools, marketed use cases cluster into three categories. Counts exceed 102 because many agents span more than one category.
2.2Usability ratings by task
Participants rated the usability of two tools — OpenAI Operator and Manus — on three representative tasks using the System Usability Scale (SUS, 0–100, higher is better).
3.Five barriers to agent delegation.
Despite generally positive usability scores, participants reported recurring frustrations. Five themes emerged from the qualitative data; click a card to reveal a representative quote.
Misaligned mental models
Users cannot predict what an agent is actually able to do.
“You've got to sit there and make sure that your initial prompt is perfect.”
Users can't predict the agent's actual capabilities, turning delegation into frustrating 'prompt gambling.'
Presuming trust
Agents expect delegation before earning user confidence.
“I was waiting for it to ask me for some more information.”
Agents expect immediate delegation without first establishing credibility through preference-elicitation or demonstrating competence for sensitive tasks.
Inflexible collaboration
Agents do not adapt their interaction style to the user or task.
“It's kind of like I'm giving you a job, and you're throwing the job back at me.”
They act as 'lone wolf' execution tools that fail to adapt to a user's need for hands-on guidance or mid-task oversight.
Communication overhead
Agents overwhelm users with output and ask for inarticulable input.
“Oh, my God! It threw out so much stuff... it's almost an overwhelming amount of information.”
Agents generate excessive, poorly-formatted output and force users to articulate complex, subjective preferences in a cognitively demanding way.
Metacognitive gaps
Agents lack self-awareness of their errors and limitations.
“It just was kind of circling... it's seeking to provide an answer rather than to say 'I don't know.'”
They lack the self-awareness to recognise their own errors, leading to time-wasting 'try-fail cycles' that require manual human debugging.
4.Six recommendations for next-generation agents.
Know your user
Actively collect preferences, skills, and collaboration styles before executing a task.
Know yourself
Develop metacognitive abilities to recognise capability limits and stop confidently.
Be adaptable
Adapt the interface and interaction style based on task type and user signals.
Measure twice, cut once
Surface planning and checkpoints so users can steer before and during execution.
Show, don't tell
Support multiple input modalities beyond free-form text prompts.
Next time's the charm
Enable precise iteration on outputs via contextual, direct-manipulation controls.
Citation.
@inproceedings{shome2026johnny,
title={Why Johnny Can't Use Agents: Industry Aspirations vs. User Realities with AI Agent Software},
author={Shome, Pradyumna and Krishnan, Sashreek and Das, Sauvik},
booktitle={Proceedings of the ACM Conference on AI and Agentic Systems (CAIS '26)},
year={2026}
}