AI agents are systems that can automatically perform tasks based on a generative model, long been described as a virtual assistant that can book tickets, schedule meetings or process repeat requests on behalf of users.
However, a large-scale study by Harvard University has just published a completely different picture of how people actually use these tools.
According to an analysis of hundreds of millions of anonymous interactions on Perplexity's Comet and Comet Assistant browsers, more than 57% of AI agents are related to cognitive tasks, not simple administrative operations.
Specifically, 36% of popular actions serve the purpose of productivity and work processes; the remaining 21% is for learning and research.
The study is considered the first major effort to measure AI actor usage habits, focusing on answering three questions, including: who uses AI assistants, how often they use them, and what tasks they assign.
The results show that productivity-related tasks also have the highest user retention rate. People who start with study or research tasks tend to continue to use them for a long time.
This comes as companies around the world are accelerating the adoption of AI. The findings of Harvard challenge the common perception that AI agents are primarily for small-scale work. Instead, users are seeing them as tools to support thinking, learning and problem solving.
Perplexity commented: Data shows that we are entering a hypersonic economy. AI workers are expanding the scale of human cognitive work, and the level of engagement will be even deeper when the tool is complete.
Although 2025 is expected to be the year of AI agents, the speed of actual application has not exploded as predicted.
Many experts, including artificial intelligence (AI) researcher Andrej Karpathy, expressed skepticism about the development roadmap: "Calling this the year of AI agents is too much. To be more accurate, it must be a decade.
The study also analyzed the professional skills of Comet users. In particular, fields such as marketing, sales, management and startups recorded the highest level of attachment to AI.
Data shows that users take advantage of AI agents to solve specific difficulties of each industry. In finance, 47% of the questions were related to productivity tasks; while students had 43% of the questions about studying and researching.
More interestingly, usage behavior often evolutionates when users start with simple questions such as suggesting movies, then gradually switch to complex requirements such as removing code errors or summarizing financial reports.
Once they have entered this task force, they rarely return to normal use.
Harvard's research opens up a new perspective that AI agents are no longer just a tool to support operations, but are becoming a companion in intellectual work, a factor that can reshape the way people learn, work and make decisions.
