Anthropic Agents Overwhelm Human Workers With Data
Anthropic's AI agents are producing vast amounts of data, overwhelming human workers and creating a 'reading tax' that threatens productivity. Designers must adapt to keep humans in charge.
Original reporting · uxdesign.cc
Original: “Human attention is the new old bottleneck. We have to design for it even more.”
Read “Human attention is the new old bottleneck. We have to design for it even more.” at UX Collective →
How this was written. Awesome B2B Design wrote this article with a language model from UX Collective’s reporting (model: @cf/meta/llama-3.3-70b-instruct-fp8-fast); it describes that reporting rather than adding to it. The original, linked above, is the full story.
Key points from UX Collective’s report
- Anthropic's AI agents are producing vast amounts of data, overwhelming human workers and creating a 'reading tax' that threatens productivity.
- Designers must prioritize human attention and create systems that filter out unnecessary information to mitigate the reading tax.
- Ranking agent requests, designing the review step, and writing decisions down are key steps in keeping humans in charge of the loop.
- The problem of information overload is not new, and researchers have been warning about it for decades.
- Designers can learn from examples in other fields, such as hospital alarm management, to create more effective systems.
Anthropic, a company at the forefront of AI development, has published a study on how its agents are changing the nature of work. According to UX Collective, the study reveals that agents are logging six hours for every human hour, producing vast amounts of data that humans must review.
The study found that agent sessions spend about 80% of their time idle, waiting for a human to respond, and that reading everything an agent produces would take more than half the workweek. This 'reading tax' is a major bottleneck in productivity, and designers must find ways to mitigate it.
The problem is not new, and researchers have been warning about the dangers of information overload for decades. Herbert Simon predicted in 1971 that 'a wealth of information creates a poverty of attention,' and Gloria Mark's studies have shown that people's attention span is already thin, with the average person spending only 47 seconds on one screen.
To address this issue, designers must prioritize human attention and create systems that filter out unnecessary information. One approach is to measure reading time and treat it like a budget that teams must set and defend. This means designing interfaces that prioritize the most important information and minimize distractions.
Another key step is to rank agent requests and prioritize the most critical ones. This can be achieved by making agents justify the severity level of each request and keeping routine ones in the periphery. Hospitals have already implemented similar systems to manage alarm fatigue, and designers can learn from these examples.
Designing the review step is also crucial, as it is where humans pay the reading tax and either keep agency or hand it over to agents. Review screens should show what changed, what the agent checked and skipped, and where each claim came from. Logging review time per item can also help teams see the tax and adjust their workflows accordingly.
Finally, writing decisions down and making them visible to agents is essential. Agents reason from whatever people leave in the open, and unclear decisions can lead to confusion and errors. By prioritizing human attention and designing systems that support it, designers can help keep humans in charge of the loop and ensure that agents augment human productivity rather than overwhelming it.
Questions and answers
What is the 'reading tax' and how does it affect productivity?
The 'reading tax' refers to the time and effort required for humans to review and process the vast amounts of data produced by AI agents, which can overwhelm human workers and threaten productivity.
How can designers mitigate the reading tax and keep humans in charge of the loop?
Designers can mitigate the reading tax by prioritizing human attention, ranking agent requests, designing the review step, and writing decisions down, among other strategies.
What can designers learn from other fields, such as hospital alarm management, to create more effective systems?
Designers can learn from hospital alarm management systems, which prioritize critical alarms and minimize distractions, to create similar systems for managing agent requests and minimizing the reading tax.
