Can autonomous web agents collaborate with humans without creating hidden work?

Do autonomous agents actually reduce your workload, or just shift it?

The short answer is that it depends entirely on how the agent is built. In a 2025 study of a human-agent collaborative web navigation system called CowPilot, humans performed only 15.2% of the total steps while the agent handled the rest, and the team achieved a 95% success rate [4]. That means the agent did the heavy lifting — filling forms, navigating pages, clicking buttons — and the human only stepped in for critical decisions or corrections. The key design feature was that the agent proposed next steps, and the human could pause, reject, or override them without having to micromanage every action [4]. This suggests that when agents are designed to be interruptible and transparent, they genuinely reduce visible work rather than creating hidden tasks.

However, a 2023 study of conversational agents in virtual workplaces found a darker side: agents can collect user data, reduce worker autonomy, and foster social isolation, all of which create hidden work in the form of privacy management, loss of control, or emotional labor [2]. The researchers interviewed 15 senior experts in ethics, collaboration, and computer science to derive 14 ethical guidelines for introducing such agents [2]. The takeaway is that hidden work isn't just about extra clicks — it can be psychological or organizational, like having to monitor what the agent does with your data or feeling pressured to keep up with its pace.

What design features prevent hidden work?

The most effective agents share two traits: they are transparent about their actions and they allow humans to easily intervene. In the CowPilot study, users could interleave their own actions with the agent's, overriding suggestions or resuming agent control at any time [4]. This flexibility meant the human never had to redo work the agent already did, nor did they have to plan every step in advance. Even when humans intervened during task execution, the agent still drove up to half of task success on its own [4].

Another approach comes from a 2024 study on hierarchical ad hoc agents (HA²) in a collaborative cooking game. The agents were designed to mimic how humans break down tasks into shared abstractions — essentially, a common understanding of what needs to happen and in what order [3]. When paired with unseen human teammates, HA² agents showed statistically significant improvement over existing methods, meaning they adapted to human behavior without requiring the human to learn a new system [3]. This reduces hidden work because the human doesn't have to explain or teach the agent how to do its part.

A 2022 study on human-agent teams in a time-pressured pursuit task found that agents with an active stake in the team's success — where the agent's own performance depended on the team outcome — actually offset loss of situation awareness in humans [5]. In plain terms, when the agent was invested in the team's goal, humans didn't have to constantly monitor what the agent was doing, which reduced the hidden cognitive load of keeping track of a teammate's actions [5].

When does collaboration backfire and create hidden work?

The evidence points to several conditions where agents can increase hidden work. First, if the agent is unreliable or unpredictable, humans must constantly double-check its actions, which defeats the purpose of automation. A 2022 study found that reducing an agent's reliability could paradoxically make it more effective if the change was imperceptible to the human, but the same study warned that overt unreliability forces humans to compensate, increasing their workload [5].

Second, agents that lack transparency about their data collection create hidden work in the form of privacy management. The 2023 ethical guidelines study explicitly warns that conversational agents can collect user data without users' full awareness, requiring humans to invest time in understanding and controlling what data is being gathered [2]. This is a form of hidden work that isn't obvious at first — you might not realize you're spending mental energy worrying about privacy until it's too late.

Third, agents that are too rigid or require extensive setup can shift the burden onto humans. For example, a 2024 study on simple autonomous agents in creative discovery tasks found that bots sharing the most similar noun improved group performance only when the semantic space was easy to navigate [1]. In harder tasks, the same bots might have confused the group, forcing humans to spend extra time interpreting or correcting the agent's suggestions [1]. The lesson is that agents designed for simple tasks can create hidden work when applied to complex, ambiguous problems.

About These Sources

This answer is built on 5 peer-reviewed studies — published from 2022 to 2025, 3 from 2024 or later, 2 in Q1 journals — selected as the most relevant from 5 studies that passed quality screening, drawn from 29 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions