The Shift from Search to Autonomous Execution
A joint research study from Harvard and Perplexity compares user behavior between a conversational search engine and an autonomous agent (Perplexity Computer). By analyzing 10,000 matched query pairs, the study highlights how agents fundamentally change the cost structure of knowledge work. While search engines excel at quick lookups (median 14 seconds), agents handle complex, multi-step workflows, performing a median of 9 minutes of autonomous machine work per session.
The Cost-Structure Framework
The research introduces a model where tasks have a fixed cost (delegation and review) and a marginal cost (execution).
- Search: Low fixed cost, high marginal cost per step ($2.05/step).
- Agent: Higher fixed cost, significantly lower marginal cost ($0.16/step).
This creates a clear breakeven point: short, simple queries remain cheaper in search, while complex, multi-step tasks are more efficient when delegated to an agent. The study estimates that for complex tasks, the combination of a human and an AI agent reduces total time by 87% and total cost by 94% compared to a human using only search.
Expanding the Scope of Work
Beyond mere speed, agents change the nature of the tasks users attempt:
- Cognitive Demand: 76% of agent queries require higher-order cognition (Bloom’s Taxonomy) compared to 55% for search.
- Domain Breadth: Agent queries are more likely to span multiple knowledge domains (2.40 domains per query vs 1.74).
- Task Complexity: Agents are used for 'Create-level' work 50% of the time, double the rate of search.
Crucially, the study finds that agent adoption is complementary rather than substitutive; users who adopt agents actually increase their daily search volume, suggesting that agents unlock new categories of work that were previously too time-consuming to attempt.