New Research from Harvard and Perplexity Finds AI Agents Perform 26 Minutes of Autonomic Activity Per Session vs 33 Seconds of Search
New research in action from Perplexity and Harvard provides field evidence on what AI agents are doing in information work. It draws on production data from two Perplexity products: Search and Computing.
The setting is a natural comparison. Search is a conversational response engine. A computer is an agent that organizes and executes end-to-end operations. The same users touch both products, so the team can handle work almost constantly.
What the Course Represents
The research study covers a 90-day window, February 27 to May 27, 2026. The computer was launched two days before that window opened.
The core method is similar to query pairs that are almost identical across the two products. The research team found 10,000 pairs of sessions with a cosine similarity greater than 0.99. Each pair is the same task attempted in both ways.
Computer pairs that are included in the gateway sessions request the startup tool. These ‘doing’ tools include code execution, browser actions, file scripting, and connector calls. That gateway ensures that every Computer session does real independent work.
Adoption went up the window. Combined Computer Quiz reached 84× its first week’s value. A similar analysis found that Computer acceptance also raised daily user search queries by 1.05. A positive effect indicates complementarity, not substitution.
Cost Structure Outline
The study bases its data on a simple task-based model. Each task has a step value, and long tasks have a weak peak value.
Agents change the cost structure. They charge a high fixed fee for each transaction, referral and review. But they charge a lower marginal cost for each step, as the system works.
This produces a step separation equation. Below it, the chat mode is cheap. Above it, the agent mode wins. Short-term residence permit; long workflow to the agent.
Autonomy: 26 Minutes vs 33 Seconds
The first measure of independence is the time to do it. The computer uses 26 minutes of machine work per session. Search takes 33 seconds. That’s a 48× gap.
The medians show the same pattern: 9 minutes versus 14 seconds. The gap varies by domain. Area functions show 75×; Science shows 26 ×, as clear answers are usually enough.
High independence did not lower the standard here. The research team found the following dissatisfaction from what users do next. Computer’s perceived dissatisfaction rate was 1.3%, compared to 2.9% for Search (a 55% reduction).
Tracking curves are also changing for revisions and expansions on PC, although the changes are minor. Connector usage increased more clearly. Computer requested at least one connector 7.9% of the time, compared to 1.8% for Search. The computer incorporates external tools that Search users would otherwise operate manually.
Efficiency: Where Savings Come From
The efficiency section measures the Search + Human counterfactual. A person with Search alone takes 269 minutes for each activity matched. Computer + Human takes 36 minutes.
That’s 87% less time and 94% lower costs. Cost savings outweigh time savings because domain salaries increase productivity. Computer model costs run $4–10 per job; Searches use about $0.05.
The numbers on the side support the frame. Computer + Human costs $0.16 per step, compared to $2.05 for Search + Human. Computational computer times also used longer information, 652 characters compared to 448 for borders. That supports the high assumption of fixed costs for agents.
Breakeven analysis states that a technician must complete all manual steps in less than 20 minutes to simulate a Computer. The research team carefully evaluated LLM’s independent rating and interviews with users. The LLM method achieved 84% time and 93% cost savings. Interviewers reported acceleration from 5× to 300×.
Horizontal and vertical expansion
Scope is where this study extends previous work. Autonomy doesn’t just speed up jobs. It changes what tasks users try.
Horizontally, computing queries often cross lines of work. The share of various jobs is around 59% in Computing, compared to 50% in Search. Management and business showed the largest gap, with 19 points.
Specifically, Computer questions are very demanding. In Bloom’s Revised Taxonomy, 76% require higher-level understanding, compared to 55% for Search. Creative level work accounted for 50% of Computing questions, compared to 26%.
Computer jobs also include additional information domains. Each query affected 2.40 O*NET databases on average, compared to 1.74. It was about three times as much to need three or more domains.
Composability increases as the O*NET class improves. At the Statement of Work level, the Computer has done 60% more work. About 23% of Computer queries reach a Job Statement that the same users never submitted to Search.
Comparison table: Search vs Computer
| Size | Anomalous Search | A Confused Computer |
|---|---|---|
| Mode in frame | Chat response engine | Orchestrator agent |
| Machine time per session | 33 seconds (median 14s) | 26 minutes (9m average) |
| Questions for each session | 2.8 | 5.3 |
| Meaningful dissatisfaction (medium+high). | 2.9% | 1.3% |
| Sessions have a connector call | 1.8% | 7.9% |
| Time for real work | 269 min (Search + Person) | 36 min (Computer + Person) |
| Cost per step | $2.05 | $0.16 |
| Model costs per job | ~$0.05 | $4–10 |
| Sharing cross-occupation questions | 50% | 59% |
| Higher-order Bloom cognition | 55% | 76% |
| O*NET Databases per query | 1.74 | 2.40 |
Key Takeaways
- The computer uses 26 minutes of independent work per session compared to 33 seconds for Search, a 48× gap.
- In the simulated tasks, Computer + Human cut estimated time by 87% and cost by 94% compared to Search + Human.
- The average dissatisfaction rate for computer is 1.3% compared to 2.9% for Search, a decrease of 55%.
- Computer queries cut across multiple tasks (59% vs 50%) and require advanced recognition (76% vs 55%).
- About 23% of Computer queries hit the Job Statement the same users never submitted to Search.
Marktechpost Visual Explainer
Research Guide
Harvard × Confusion
Marktechpost
Cover of early AI/ML research, recorded by developers.
Check it out Paper again Technical details. Also, feel free to follow us Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.
Need to work with us on developing your GitHub Repo OR Hug Face Page OR Product Release OR Webinar etc.? contact us
