When does hidden work emerge in human-multi-agent teams?
Hidden work—unseen effort like correcting agent mistakes, deciphering opaque outputs, or managing unexpected agent behaviors—arises when agents operate without transparency or human oversight. A 2023 interview study with 15 senior researchers in ethics, collaboration, and computer science identified that conversational agents can collect user data, reduce worker autonomy, and foster social isolation, all of which shift invisible burdens onto humans [2]. Similarly, a 2026 exploratory study with 12 design practitioners found that when humans formed teams with multiple AI agents for creative work, participants initially tried to let agents operate autonomously but ultimately had to directly orchestrate them because inter-agent interactions created complexity and unexpected behaviors [5]. This means that without explicit design for human control, agents generate hidden work by requiring constant supervision and correction.
What design features prevent hidden work?
Two key design strategies—incremental collaboration and transparent reasoning—can eliminate hidden work. A 2026 study introduced MAVIS, a multi-agent system that automatically decomposes tasks into guidelines and sequentially introduces expert agents, each proactively engaging users in focused dialogue to uncover implicit preferences. This incremental mechanism doubled the number of user preferences expressed and improved planning quality by 60.3% over a conventional LLM baseline, meaning users spent less effort correcting or re-explaining their needs [1]. MAVIS also used spatial visualizations—step-linked summaries and context-aware boards—to externalize agents' reasoning, reducing the cognitive load of tracking what each agent was doing [1]. A separate 2026 framework for industrial multi-agent systems proposed using a Digital Twin as a service layer for system coordination, supervision, and interaction, creating a transparent and seamless collaboration between agents and humans [4]. Both studies converge on the same conclusion: making agent reasoning visible and giving humans control over when and how agents act prevents hidden work.
Are there limits to preventing hidden work?
Even with transparent design, hidden work can persist depending on the task and the number of agents. The 2026 creative-work study found that as the number of AI agents increased, inter-agent interactions became harder to predict, forcing humans to step in and orchestrate the team directly—a form of hidden work [5]. This suggests that for complex, open-ended creative tasks, full autonomy may be impractical, and some level of human orchestration is unavoidable. Additionally, the industrial Digital Twin framework, while promising, was only evaluated in a single real-world case study and requires long-term evaluations across different contexts to confirm it consistently prevents hidden work [4]. The 2021 task-scheduling study in manufacturing used deep multi-agent reinforcement learning to optimize human-robot collaboration without expert knowledge, but it focused on efficiency (completion time) rather than transparency or hidden work, leaving open the question of whether such optimization creates unseen burdens for human workers [3]. So while transparency features can dramatically reduce hidden work, they may not eliminate it entirely for highly complex or multi-agent scenarios.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2021 to 2026, 3 from 2024 or later, 2 in Q1 journals, collectively cited 125 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 44 papers retrieved from a database of over 500 million.
Sources used in this answer
1
Virtual Minds, Real Work: LLM-Powered Preference-Based Planning through Spatial Multi-Agent-Human Collaboration.
In a controlled study, the MAVIS multi-agent system used incremental collaboration and spatial visualizations to double expressed user preferences and improve planning quality by 60.3% over a conventional LLM baseline, reducing hidden work by making agent reasoning transparent.
2
Ethical Challenges for Human–Agent Interaction in Virtual Collaboration at Work
Through 15 expert interviews, this 2023 study identified that conversational agents can create hidden work by collecting user data, reducing worker autonomy, and fostering social isolation, leading to 14 ethical guidelines for safer human-agent collaboration.
3
Optimizing task scheduling in human-robot collaboration with deep multi-agent reinforcement learning
This 2021 study applied deep multi-agent reinforcement learning to optimize task scheduling in human-robot collaboration, achieving efficient real-time decisions without expert knowledge, but did not address transparency or hidden work burdens.
4
From Automation to Autonomy: A Digital Twin Framework for Transparent Agent and Human Collaboration in Industrial Multi-Agent Systems
This 2026 study proposed a Digital Twin framework for transparent human-agent collaboration in industrial multi-agent systems, demonstrated in a real-world case study to provide a solid technical foundation for seamless, visible coordination.
5
Understanding Human-Multi-Agent Team Formation for Creative Work
In a 2026 exploratory study with 12 design practitioners, humans forming teams with multiple AI agents for creative work initially attempted autonomous operation but ultimately had to directly orchestrate agents due to unexpected inter-agent behaviors, revealing hidden work in complex multi-agent setups.
Related research questions
