A new study conducted by researchers from Harvard and Perplexity shows that AI agents significantly outperform traditional search tools, executing an average of 26 minutes of autonomous work per session, compared to just 33 seconds for search functions. This finding highlights the changing dynamics of knowledge work and the capabilities of AI-driven tools.
Study Overview
The research, conducted over 90 days from February 27 to May 27, 2026, analyzed data from two Perplexity products: Search, a conversational answer engine, and Computer, an agent designed to manage tasks from start to finish. By using a natural comparison framework, the study allowed users to interact with both products while performing similar tasks, leading to a more accurate assessment of their performance.
To establish reliable metrics, the research team matched query pairs across the two platforms. They identified 10,000 session pairs with a cosine similarity of 0.99, ensuring that each pair represented nearly identical tasks. Sessions with the Computer agent were limited to those involving execution tools, such as code execution and file writes, confirming that the measured work was genuinely autonomous.
Key Findings
The results revealed a remarkable increase in the adoption of the Computer tool throughout the study period. Cumulative queries for Computer surged by 84 times compared to its initial week. Users who engaged with the Computer agent also increased their daily Search queries by an average of 1.05, indicating a complementary rather than a substitutive relationship.
The study utilized a task-based model to evaluate performance, assigning a step count to each task. This model showed that longer tasks tend to provide higher value. Additionally, the research pointed out a shift in cost structures following the introduction of AI agents. While agents incur higher fixed costs for task delegation and review, they lower the marginal cost per step, leading to a breakeven point where using an agent becomes more economical for longer workflows.
Implications for Knowledge Work
These findings highlight a significant shift in how knowledge work is conducted. Shorter tasks still benefit from traditional search methods, as they are more cost-effective in those situations. However, as task complexity and length increase, the advantages of AI agents become clear. This shift suggests that as organizations adopt these tools, the nature of work will evolve, enabling greater efficiency in managing complex workflows.
As businesses continue to explore the integration of AI agents into their operations, the potential impact on productivity and task management could be substantial. The research indicates a future where AI-driven tools not only assist in knowledge work but also redefine efficiency and execution standards.
Quick answers
What was the main finding of the study?
AI agents perform an average of 26 minutes of autonomous work per session, compared to 33 seconds for search tools.
How did the researchers ensure the tasks were comparable?
The study matched 10,000 query pairs across both products with a cosine similarity of 0.99.
What was the impact of using the Computer tool on search queries?
Users increased their daily Search queries by an average of 1.05 when using the Computer agent.
What is the breakeven point in terms of task length?
The breakeven point occurs where longer tasks become more cost-effective with AI agents compared to traditional search.
The stories that move AI & crypto markets — before the market reacts.
Free. 7am ET. Five stories. 62,400 readers.
