A New Study from Harvard and Perplexity Finds AI Agents Perform 26 Minutes of Autonomous Work per Session vs 33 Seconds for Search
A new working analysis from Perplexity and Harvard gives subject proof on what AI brokers do to data work. It attracts on manufacturing knowledge from two Perplexity merchandise: Search and Computer.
The setup is a pure comparability. Search is a conversational reply engine. Computer is an agent that plans and executes duties finish to finish. The identical customers contact each merchandise, so the workforce can maintain the duty roughly fixed.
What the Study Actually Measures
The analysis examine covers a 90-day window, February 27 by May 27, 2026. Computer launched two days earlier than that window opened.
The core technique matches near-identical question pairs throughout the 2 merchandise. The analysis workforce discovered 10,000 session pairs with cosine similarity above 0.99. Each pair is successfully the identical activity tried each methods.
Computer pairs are gated to periods that invoke an execution device. These ‘do’ instruments embrace code execution, browser actions, file writes, and connector calls. That gate ensures each Computer session does actual autonomous work.
Adoption rose over the window. Cumulative Computer queries reached 84× their first-week whole. A matched evaluation discovered Computer adoption additionally raised customers’ each day Search queries by 1.05. The optimistic impact factors to complementarity, not substitution.
The Cost-Structure Framework
The analysis grounds its knowledge in a easy task-based mannequin. Each activity has a step depend, and longer duties carry weakly larger worth.
Agents change the fee construction. They cost the next mounted value per activity, for delegation and evaluate. But they cost a decrease marginal value per step, because the system executes.
This produces a breakeven step depend. Below it, the conversational mode is cheaper. Above it, the agent mode wins. Short lookups keep handbook; lengthy workflows transfer to the agent.
Autonomy: 26 Minutes vs 33 Seconds
The first autonomy measure is execution time. Computer runs 26 minutes of machine work per session. Search runs 33 seconds. That is a 48× hole.
Medians present the identical sample: 9 minutes versus 14 seconds. The hole varies by area. Local duties present 75×; Science exhibits 26×, since plain solutions typically suffice.
Higher autonomy didn’t decrease high quality right here. The analysis workforce scored next-turn dissatisfaction from what customers do subsequent. Computer’s significant dissatisfaction charge was 1.3%, in opposition to 2.9% for Search (55% discount).
Follow-up turns additionally shift towards evaluate and extension on Computer, although the adjustments are small. Connector utilization rose extra clearly. Computer invoked a minimum of one connector in 7.9% of periods, versus 1.8% for Search. Computer chains exterior instruments that Search customers would in any other case run by hand.
Efficiency: Where the Savings Come From
The effectivity part estimates a Search + Human counterfactual. A human with Search alone takes 269 minutes per matched activity. Computer + Human takes 36 minutes.
That is 87% much less time and 94% much less value total. Cost financial savings exceed time financial savings as a result of area wages amplify the impact. Computer’s mannequin value runs $4–10 per activity; Search runs about $0.05.
The marginal numbers help the framework. Computer + Human prices $0.16 per step, versus $2.05 for Search + Human. Matched Computer periods additionally ran longer prompts, 652 versus 448 characters on the median. That helps the upper fixed-cost assumption for brokers.
Breakeven evaluation says knowledgeable should end all handbook steps in beneath 20 minutes to match Computer. The analysis workforce cross-checked with an unbiased LLM estimate and consumer interviews. The LLM technique discovered 84% time and 93% value financial savings. Interviewees reported speedups from 5× to 300×.
Horizontal and Vertical Expansion
Scope is the place this analysis extends previous prior work. Autonomy doesn’t simply velocity up duties. It adjustments which duties customers try.
Horizontally, Computer queries cross occupational traces extra typically. Cross-occupation share averaged 59% on Computer, versus 50% on Search. Management and Entrepreneurship confirmed the biggest hole, at 19 factors.
Vertically, Computer queries are extra demanding. On Bloom’s Revised Taxonomy, 76% required higher-order cognition, versus 55% for Search. Create-level work was 50% of Computer queries, in opposition to 26%.
Computer duties additionally span extra data domains. Each question touched 2.40 O*NET Knowledge domains on common, versus 1.74. It was almost 3 times as more likely to want three or extra domains.
Composability climbs because the O*NET hierarchy will get finer. At the Task Statement stage, Computer engaged 60% extra actions. About 23% of Computer queries hit a Task Statement that the identical customers by no means despatched to Search.
Comparison Table: Search vs Computer
| Dimension | Perplexity Search | Perplexity Computer |
|---|---|---|
| Mode within the framework | Conversational reply engine | Agent orchestrator |
| Machine time per session | 33 seconds (median 14s) | 26 minutes (median 9m) |
| Queries per session | 2.8 | 5.3 |
| Meaningful (mid+excessive) dissatisfaction | 2.9% | 1.3% |
| Sessions with a connector name | 1.8% | 7.9% |
| Counterfactual activity time | 269 min (Search + Human) | 36 min (Computer + Human) |
| Cost per step | $2.05 | $0.16 |
| Model value per activity | ~$0.05 | $4–10 |
| Cross-occupation question share | 50% | 59% |
| Higher-order Bloom cognition | 55% | 76% |
| O*NET Knowledge domains per question | 1.74 | 2.40 |
(*33*)Key Takeaways
- Computer runs 26 minutes of autonomous work per session versus 33 seconds for Search, a 48× hole.
- On matched duties, Computer + Human cuts estimated time 87% and value 94% versus Search + Human.
- Computer’s significant dissatisfaction charge is 1.3% versus 2.9% for Search, a 55% discount.
- Computer queries cross occupations extra (59% vs 50%) and demand extra higher-order cognition (76% vs 55%).
- About 23% of Computer queries hit a Task Statement the identical customers by no means despatched to Search.
Marktechpost’s Visual Explainer
Check out the Paper and Technical details. Also, be happy to observe us on Twitter and don’t overlook to affix our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to companion with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and so on.? Connect with us
The put up A New Study from Harvard and Perplexity Finds AI Agents Perform 26 Minutes of Autonomous Work per Session vs 33 Seconds for Search appeared first on MarkTechPost.
