Autonomous AI agents run 50x longer than search: Harvard and Perplexity research

Researchers from Harvard University and the company Perplexity published on June 8, 2026 a joint work that directly compares user behavior when working with an autonomous AI agent and a classical search-based AI assistant. The main finding: on average, a session with an autonomous agent takes 26 minutes of independent system operation, whereas a session with a search assistant takes only 33 seconds. The difference is almost 50 times in favor of the autonomous agent.

How the study was structured

The authors applied a matched-pair sessions method: they compared tasks similar in nature that users solved either through an autonomous agent or through a search service, in order to correctly compare human behavior in two different modes of working with AI. This experimental design allows separating the effect of the tool itself from the effect of the type of task the user is solving — that is, the comparison turns out to be honest, not biased in favor of one of the systems due to differences in request complexity. The work continues a broader line of research in 2026 devoted to how to measure the "task horizon" — that is, how many sequential steps a system can perform independently before requiring human intervention.

It is the growth of this indicator, and not just the improvement in the quality of a single response, that analysts increasingly call the main indicator of progress in AI agent development.

Key figures from the study:

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

What the figures showed, beyond time

The difference in session time is not the only result. Researchers record broader gains in three directions: the degree of autonomy (the agent performs more sequential steps without intermediate confirmation from the user), the total time a person saves by delegating a task entirely, and the cost of performing the task in terms of the result achieved. Additionally, the authors note that when working with an autonomous agent, users generally take on a wider range of tasks than when using search — that is, the agent not only accelerates familiar information search scenarios, but expands the very list of what 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 ratio of 26 minutes versus 33 seconds clearly explains why the industry in 2026 is so rapidly shifting focus from chatbots and search assistants to full-fledged agents capable of bringing multi-step tasks to completion independently. For Perplexity, which is simultaneously developing both search products and agent functions, this data is a weighty argument in favor of further investment in autonomy, not just in the quality of a single response to a query.

For the AI market as a whole, the result confirms a thesis that the industry has been discussing for months: the value of AI is shifting from "quick answer" to "work completed." If an agent really occupies a user with almost 50 times more independent work without human intervention, this means exponential potential growth in the volume of tasks that businesses and individual users are willing to delegate to the system. For investors and product teams, such figures also become a guide when choosing where to direct resources — toward improving search or toward developing full-scale agent autonomy.

The question of quality and reliability of this work on long horizons remains open — this, it seems, will become the next focus of Harvard and Perplexity research in this area.

It is also important that the comparison is built on actual user sessions rather than synthetic benchmarks, which the industry traditionally uses to measure model capabilities. This approach eliminates some of the criticism usually directed at laboratory tests of agents: real users set the system more diverse and less predictable tasks than standardized test sets, which means the gap of 26 minutes versus 33 seconds is closer to what happens in practice, not to laboratory maximum model capabilities.