AI agents are rewriting how knowledge work gets done. They now run 26 minutes of autonomous tasks per session—compared to mere seconds for traditional search.

A recent study tracked user sessions on two Perplexity products: Search, a conversational answer engine, and Computer, an autonomous agent that plans and executes tasks end to end. The same users performed nearly identical queries on both platforms. The result? Autonomous agents handle workflows with 48 times more machine execution per session.

Search sessions lasted about 33 seconds of machine time. Computer sessions clocked in at 26 minutes. Medians show similar differences: 14 seconds versus 9 minutes. The gap is bigger in local tasks and smaller in straightforward scientific lookups, where simple answers often suffice.

The autonomy boost comes with quality gains. User dissatisfaction dropped by more than half when agents handled tasks, signaling fewer dead ends and better results. Agents also chained external tools more often, automating what users would otherwise do manually.

Efficiency gains are staggering. The study estimates that humans using search alone spend 269 minutes per task. When paired with the agent, that drops to 36 minutes—saving 87% time and 94% cost. Agents carry a higher fixed cost but slash marginal costs per workflow step. They pay off for longer, complex tasks where delegation beats manual effort.

Search as Code: AI Writing Its Own Pipelines

Perplexity’s latest innovation lets AI models write their own search pipelines as Python code. Traditional search APIs offer fixed endpoints—models ask a query, get results, and refine. This loop suits humans, not autonomous agents that run hundreds of searches rapidly.

Search as Code flips that model. The AI generates custom search workflows to retrieve, filter, deduplicate, and rerank results by composing atomic primitives exposed via a new SDK. The code runs in a secure sandbox with persistent state, enabling complex, multi-step research.

This approach cuts token consumption by 85% and hit 100% accuracy on a tough software vulnerability detection benchmark, outperforming OpenAI and Anthropic. It shifts bottlenecks from data retrieval to strategy design—favoring models strong in reasoning and code generation.

Multi-Model Orchestration Powers Complex Workflows

Perplexity’s Computer agent coordinates 19 AI models dynamically, each specialized in subtasks like deep research, coding, summarization, or image generation. The system decomposes user goals, routes subtasks, and manages asynchronous workflows that can run for hours or months without human oversight.

Claude Opus 4.6 handles orchestration and heavy reasoning. Google Gemini manages deep research. GPT-5.2 sustains long-context recall. Lightweight tasks go to Grok. This modular approach beats betting on a single model and lets Perplexity swap models as better ones emerge.

Cloud execution offers isolation and compliance for enterprises. The architecture supports over 400 app connectors and integrates with corporate data platforms, aiming at high-stakes tasks like competitive intelligence, due diligence, and strategic analysis.

Perplexity’s $200/month Computer subscription targets professionals. It’s not a casual chatbot but a productivity accelerator designed for complex, multi-step workflows that demand auditability and reliability.

Unlike OpenAI and Anthropic, which focus on singular flagship models, Perplexity builds a multi-model platform. It routes tasks to the best model rather than forcing users into one ecosystem. This strategy appeals to enterprises that prioritize trust, transparency, and accountability.

Perplexity’s architecture pays off in real-world usage. Users rely on AI for everyday cognitive work—document editing, data filtering, scheduling—not just imaginative tasks. The platform reduces hallucinations by cross-checking multiple models and grounding answers in live web data with citations.

In the emerging AI agent economy, Perplexity positions itself as the operating system, not just another AI brain. If the market grows as predicted, the biggest winner might be whoever orchestrates the models, not the model creators themselves.