Autonomous AI agents run about 50 times longer than search assistants, Harvard and Perplexity study finds

Researchers at Harvard University and Perplexity published a joint study on June 8, 2026 that directly compares how users behave with an autonomous AI agent and with a classic AI search assistant. The main finding: an average session with an autonomous agent involves 26 minutes of independent work by the system, while a session with a search assistant lasts only 33 seconds. That is a difference of almost 50 times in favor of the autonomous agent.

How the study was designed

The authors used matched-pair sessions: they paired tasks of similar nature that users solved either through an autonomous agent or through a search service, so they could fairly compare behavior in two different modes of working with AI. This design separates the effect of the tool itself from the effect of the type of task the user is solving. The comparison is therefore fair, not skewed in favor of one system by differences in query complexity.

The work continues a broader line of 2026 research on how to measure the "task horizon": how many sequential steps a system can complete on its own before it needs human intervention. Analysts increasingly name growth in this metric, not just better single answers, as the main indicator of progress in AI agent development.

Key figures from the study:

• 26 minutes: average autonomous agent working time per session
• 33 seconds: average working time of an AI search assistant per session
• Difference: almost 50 times in favor of autonomous mode
• Authors: a joint team from Harvard University and Perplexity
• Publication date: June 8, 2026

What the numbers show besides time

The difference in session length is not the only result. The researchers record broader gains in three areas: degree of autonomy (the agent performs more sequential steps without intermediate confirmation from the user), total time the person saves by delegating a task entirely, and the cost of completing a task relative to the result achieved. The authors also note that with an autonomous agent, users take on a broader range of tasks than they do with search. The agent does not merely speed up familiar information-search scenarios; it expands the list of things people try to do with AI, trusting it with more complex and longer chains of actions.

What this means for the future of AI agents

The 26 minutes versus 33 seconds ratio clearly explains why the industry in 2026 is shifting its focus so quickly from chatbots and search assistants to full agents able to carry multistep tasks through to the end on their own. For Perplexity, which develops both a search product and agent features, these data are a weighty argument for further investment in autonomy, not just in the quality of a single answer to a query.

For the AI market as a whole, the result supports a thesis the industry has been discussing for more than a few months: the value of AI is shifting from the "quick answer" to the "completed work." If an agent really occupies a user's task for almost 50 times longer with independent work and no human intervention, that implies a multiple potential increase in the volume of tasks that businesses and individuals are willing to delegate to the system. For investors and product teams, such figures also serve as a guide on where to put resources: improving search or developing full agentic autonomy.

The question of the quality and reliability of this work over long horizons remains open, and that, judging by everything, will be the next focus of Harvard and Perplexity research in this area.

It also matters that the comparison is built on real user sessions, not on the synthetic benchmarks the industry traditionally uses to measure model capabilities. This approach answers part of the criticism usually aimed at lab tests of agents: real users give the system more varied and less predictable tasks than standardized test sets. The 26 minutes versus 33 seconds gap is therefore closer to what happens in practice than to the lab maximum of a model's capabilities.