A groundbreaking study from Harvard and Perplexity reveals that autonomous AI agents fundamentally change how knowledge work gets done, completing tasks in 36 minutes that would take humans 269 minutes using traditional search alone. The research, which analyzed 10,000 matched task pairs over 90 days, shows that the gap between conversational search and autonomous agents isn't just about speed; it's about the kinds of work people attempt in the first place.

How Much Faster Are AI Agents Than Search?

The numbers are striking. Perplexity's Computer agent runs 26 minutes of machine work per session, compared to just 33 seconds for Perplexity Search, a 48-fold difference. Even at the median, the gap holds: 9 minutes for Computer versus 14 seconds for Search. The execution time varies by domain; local tasks show a 75-fold gap, while science-related queries show a 26-fold gap, since straightforward answers often suffice for factual lookups.

When researchers modeled a realistic counterfactual, the efficiency gains became even more dramatic. A professional using Search alone would spend 269 minutes on a matched task; the same person using Computer with human oversight takes just 36 minutes. That translates to 87% less time and 94% less cost overall. The cost per step tells the story: Computer plus human costs $0.16 per step, versus $2.05 for Search plus human.

Does Faster Mean Lower Quality?

One concern with delegating work to autonomous agents is whether quality suffers. The study measured user satisfaction by tracking what people did next in their sessions. Computer's meaningful dissatisfaction rate was 1.3%, compared to 2.9% for Search, a 55% reduction. Follow-up behavior also shifted; users were more likely to review and extend Computer's work rather than abandon it, suggesting they trusted the agent's output.

Connector usage provides another quality signal. Computer invoked external tools like APIs and integrations in 7.9% of sessions, versus 1.8% for Search. This matters because it shows Computer was chaining together multiple data sources and systems that Search users would otherwise have to run manually.

What Kinds of Tasks Move to Agents?

The sharpest finding from the research isn't just that agents are faster; it's that they unlock different kinds of work. Users attempted more complex, cross-disciplinary tasks on Computer than on Search. The study measured this across three dimensions:

• Cross-occupational reach: Computer queries crossed occupational lines 59% of the time, versus 50% for Search. Management and entrepreneurship queries showed the largest gap, at 19 percentage points higher on Computer.
• Cognitive complexity: On Bloom's Revised Taxonomy, 76% of Computer queries required higher-order thinking (analysis, synthesis, evaluation, creation), compared to 55% for Search. Create-level work made up 50% of Computer queries but only 26% of Search queries.
• Knowledge breadth: Computer queries touched an average of 2.40 different knowledge domains per query, versus 1.74 for Search. Computer was nearly three times as likely to require knowledge from three or more domains simultaneously.

Perhaps most striking, about 23% of Computer queries hit a task statement that the same users had never sent to Search before. This suggests that autonomous agents don't just speed up existing work; they enable entirely new categories of tasks that users previously wouldn't have attempted.

How to Choose Between Search and Agents?

The research grounds its findings in a simple economic model: agents charge a higher fixed cost per task (for delegation and review) but a lower marginal cost per step (since the system executes). This creates a breakeven point. Below it, conversational search is cheaper; above it, agents win. The study found that a professional must be able to complete all manual steps in under 20 minutes to match the agent's efficiency.

• Short lookups: Factual questions, quick reference checks, and simple searches remain cheaper and faster with traditional search. These typically involve single-step answers that don't require tool execution.
• Multi-step workflows: Tasks requiring code execution, file writes, browser automation, or external API calls benefit from agent delegation. Software engineers, data scientists, and AI engineers see the largest gains.
• Cross-domain synthesis: Work that spans multiple knowledge areas, requires higher-order thinking, or involves creating new artifacts (code, reports, visualizations) shifts decisively toward agents.

The practical lesson for organizations is task-tool fit: match the tool to the step count and complexity level. Software engineers, for example, benefit from agents that write files, run code, and deploy changes while humans supervise. Data scientists can route short lookups to conversational search but send longer analytical workflows to agents.

What Does This Mean for the Future of Work?

The study also found that Computer adoption raised users' daily Search queries by 1.05, suggesting complementarity rather than substitution. People aren't abandoning search; they're using both tools for different purposes. This aligns with how professionals have historically layered tools: a spreadsheet doesn't replace a calculator, it augments it.

The research was conducted over a 90-day window from February 27 through May 27, 2026, using production data from Perplexity's two products. The team matched nearly identical query pairs across both platforms using cosine similarity above 0.99, ensuring they were comparing the same tasks attempted two different ways. Computer sessions were gated to include only those that invoked execution tools like code runners, browser actions, file writes, and connector calls, ensuring every Computer session performed real autonomous work.

As autonomous agents mature, the question isn't whether they'll replace search; it's how organizations will integrate them into workflows where they excel. The Harvard-Perplexity findings suggest that the biggest gains come not from speed alone, but from enabling professionals to attempt work they previously couldn't or wouldn't have tried.