A recent study highlights the impact of autonomous agents on knowledge work, revealing improvements in both efficiency and quality. An arXiv paper by Jeremy Yang et al. shows that agents used in the Perplexity Computer product perform an average of 26 minutes of autonomous work per user session, compared to just 33 seconds for the Search product. The research indicates these agents not only boost productivity but also lower user dissatisfaction, with dissatisfaction rates 55% lower for Computer than Search.
The study analyzed production logs from both products and found task completion times plummeted from 269 minutes to just 36 minutes when using autonomous agents. This results in estimated time and cost reductions of 87% and 94%, respectively, compared to traditional human-led workflows. Such efficiency gains highlight the potential of these systems to manage complex, multi-step tasks that typically require manual oversight in conventional settings.
The authors noted that using autonomous agents shifts user interactions toward higher-order cognitive tasks, such as verifying and extending information. They also allow users to bundle subtasks that traditional search tools rarely accommodate. This evolution in task execution marks a significant shift in how knowledge work is conducted, transitioning from basic query responses to comprehensive problem-solving frameworks.
Technical Insights
The methodology in the study involved matched-session experiments, where nearly identical initial queries were directed to both the Computer and Search products to ensure consistency in task intent. The results strongly support the effectiveness of automated task decomposition and execution, indicating that user activities are evolving from simple query responses to more complex verification processes.
As organizations explore the potential of agentic systems, these findings provide an important empirical case study, demonstrating the efficiency gains possible when autonomy is applied to multi-step, interdependent subtasks. With user engagement shifting toward verification and cross-domain synthesis, questions arise about agent reliability, the creation of verification pipelines, and the development of evaluation metrics tailored for composite tasks.
Broader Implications
The notable reductions in time and cost—87% and 94%—underscore the importance of this research for industry practitioners aiming to enhance productivity through agentic systems. If these findings are replicated across various platforms and sectors, they could fundamentally change how tools are prioritized, benchmarks set, and engineering resources allocated.
As the trend toward higher-order follow-ups strengthens, the demand for evaluation frameworks will grow. These frameworks should not only measure accuracy but also consider verification workloads and potential failure modes. This comprehensive approach will help organizations maximize the benefits of autonomous agents while managing risks linked to automated processes.
Future Directions
Looking ahead, industry observers should monitor independent replications of these findings across different platforms and domains. Developing agent verification strategies and transparent assessments of cost savings—covering compute and supervision—will be essential for grasping the full impact of agentic systems. Additionally, research into user trust, hallucination rates during complex task execution, and the effects of crossing occupational boundaries on regulatory compliance will be vital areas of focus.
The empirical evidence presented in this study illustrates the significant efficiency and quality improvements possible through the use of autonomous agents in knowledge work, marking a critical moment for the future of productivity in this field.
The stories that move AI & crypto markets — before the market reacts.
Free. 7am ET. Five stories. 62,400 readers.