🤯 AI Agents: Revolutionizing Work & Saving You 💸

THE RISE OF AI AGENTS IN KNOWLEDGE WORK
The Perplexity and Harvard research offers compelling field evidence on how AI agents are reshaping knowledge work. This study leverages production data from two Perplexity products – Search and Computer – providing a natural comparison of their capabilities. The research focused on a 90-day period from February 27 to May 27, 2026, coinciding with the launch of the Computer agent.

COMPETITIVE ANALYSIS: SEARCH VS. COMPUTER
The research centered on comparing Search, a conversational answer engine, and Computer, an agent designed to plan and execute tasks end-to-end. The team utilized the same users for both products, maintaining a consistent task baseline. Within the 90-day window, Computer demonstrated significant growth, accumulating 84 times its initial weekly query volume. Notably, Computer adoption also stimulated users' daily Search queries by 1.05, indicating a complementary, rather than substitutional, relationship between the two tools.

TASK-BASED MODEL AND COST OPTIMIZATION
The research grounded its analysis in a simple task-based model, incorporating step counts and recognizing that longer tasks carried a higher fixed cost. Agents, like Computer, introduced a fixed cost per task alongside a lower marginal cost per step due to automated execution. This dynamic resulted in a breakeven step count, where the conversational mode became more cost-effective, and the agent mode prevailed for longer workflows. Short lookups remained manual, while extended workflows transitioned to the agent.

MEASURING AUTONOMY: EXECUTION TIME
A key measure of autonomy was execution time. Computer processed 26 minutes of machine work per session, compared to Search’s 33 seconds – a 48-fold difference. This disparity mirrored a median gap of 9 minutes versus 14 seconds, highlighting the agent's ability to automate more complex tasks. The gap varied across domains, with local tasks exhibiting the most pronounced difference (75x) and science showing a 26x reduction due to readily available answers.

REDUCED DISSATISFACTION AND TOOL UTILIZATION
The research team assessed user satisfaction through next-turn dissatisfaction rates. Computer achieved a 1.3% dissatisfaction rate, a substantial 55% reduction compared to Search’s 2.9%. Furthermore, Computer users exhibited a tendency to engage in review and extension activities, though these shifts were relatively modest. Notably, Connector usage rose more prominently with Computer, as it invoked external tools that Search users would have handled manually.

EFFICIENT TASK COMPLETION: HUMAN-AI COLLABORATION
The efficiency analysis estimated a Search + Human counterfactual, revealing a stark comparison. A human using Search alone took 269 minutes per matched task, while Computer + Human took only 36 minutes – representing an 87% reduction in time and 94% reduction in cost. Cost savings exceeded time savings due to the amplification effect of domain wages. Computer’s model cost ranged from $4 to $10 per task, while Search cost approximately $0.05 per task.

PROMPT LENGTH AND TASK COMPLEXITY
Computer sessions exhibited longer prompts, averaging 652 characters, compared to Search’s 448 characters. This difference supported the assumption of a higher fixed cost for agent-based workflows. The breakeven analysis determined that a professional must complete manual steps in under 20 minutes to match Computer’s efficiency.

VALIDATION AND EXTENT OF FINDINGS
The research team validated their findings through cross-checking with an independent LLM estimate and user interviews. The LLM method yielded 84% time and 93% cost savings, while interviewees reported speedups ranging from 5x to 300x. This demonstrated the robustness of the research’s conclusions.

SCOPE AND INNOVATION: BEYOND SIMPLICITY
The research extended beyond previous work by demonstrating that autonomy doesn’t just speed up tasks; it also influences which tasks users attempt. Horizontally, Computer queries crossed occupational lines more frequently, with an average share of 59% on Computer versus 50% on Search, particularly in Management and Entrepreneurship. Vertically, Computer queries demanded higher-order cognition, accounting for 76% of queries at the higher-order Bloom’s Taxonomy level, compared to 55% for Search. Computer tasks also spanned more knowledge domains, with an average of 2.40 ONET Knowledge domains, versus 1.74 for Search.

COMPOSABILITY AND TASK STATEMENT ENGAGEMENT
The research highlighted the concept of composability, where the ONET hierarchy's finer granularity enabled Computer to engage with 60% more activities. Approximately 23% of Computer queries involved Task Statements that the same users never sent to Search. This demonstrated the agent’s ability to handle more complex, multi-step workflows.

FURTHER EXPLORATION AND RESOURCES
Detailed information on the research, including technical specifications and a demo, can be found in the accompanying Paper and Technical details. Resources are also available via a short demo of the research paper, and a Twitter feed for updates. Joining the ML SubReddit (150k+ members) and subscribing to the Newsletter are also encouraged. Finally, a Telegram channel is available for real-time engagement and collaboration.

PARTNERING OPPORTUNITIES
Interested parties seeking to promote their GitHub Repo, Hugging Face Page, Product Release, or Webinar are invited to connect with the team.