How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and ScopeCorrespondence to Jeremy Yang ([email protected]) and Jerry Ma ([email protected]).
Exploring AI agents' impact on knowledge work, autonomy, efficiency, and scope with a focus on Perplexity's Search and Computer products.
Script: Llama 4 Scout Voice: Hume Octave 2
Transcript
Justy So, I've been reading about how AI agents are changing the game in knowledge work. Apparently, they're moving beyond just being conversational assistants to actually executing tasks on their own.
Cody Yeah, I've seen that. It's like the difference between a chatbot and a copilot. Chatbots just give you info, while copilots help you do stuff within tools. But agents take it further by connecting across tools and returning finished work with little human input.
Justy Right. And this paper from Perplexity looks at how this transition affects knowledge work using their Search and Computer products. They found that Computer, which is more autonomous, performs 26 minutes of work per session versus 33 seconds for Search.
Cody That's huge. It also says Computer reduces completion time by 87% and cost by 94% compared to humans using Search alone. Plus, it changes the scope of work users attempt, like doing more complex tasks that cross occupational boundaries.
Justy I think what's exciting is that agents like Computer can automate task decomposition and execution. This means users can focus on higher-order work like verification and extension instead of manual orchestration.
Cody Absolutely. And it's not just about efficiency; it's also about quality. The paper mentions that per-query dissatisfaction rates are 55% lower on Computer than on Search. That's a significant improvement.
Justy So, who builds with this tech? Is it mostly researchers or are companies already using it in production?
Cody From what I understand, Perplexity is already using it in their Computer product. I'd imagine other companies are exploring similar tech, but I don't have specific names.
Justy That makes sense. I could see this being really useful for knowledge workers who do repetitive or time-consuming tasks. Anyway, how do you think this tech could be built out further?
Cody Well, I'd love to see more work on how to integrate these agents with existing workflows without disrupting them. Also, ensuring they're explainable and trustworthy would be crucial for wider adoption.
Justy Yeah, we don't want rogue agents making decisions without us understanding how they got there.
Cody Exactly. But overall, I think this tech has huge potential to reshape knowledge work and make it more efficient and effective.