Developer Tools

Perplexity announces "Computer," an AI agent that assigns work to other AI agents

New tool uses Claude Opus 4.6 for reasoning, assigns tasks to specialized AI agents for hours-long projects.

Deep Dive

Perplexity has introduced 'Computer,' a sophisticated AI agent system that functions as a workflow orchestrator, assigning complex tasks to multiple specialized AI models. Currently available to Perplexity Max subscribers, Computer allows users to describe desired outcomes—such as planning a digital marketing campaign or building a custom Android app—and then autonomously breaks these into subtasks, delegating them to what Perplexity deems the best model for each job. This represents a shift from single-model agents like Claude Cowork, embracing a multi-model, best-tool-for-the-task philosophy. The system is designed to run for extended periods, from hours to potentially months, operating entirely in the cloud within isolated compute environments that have access to a real filesystem, browser, and curated tool integrations.

Technically, Computer uses Anthropic's Claude Opus 4.6 as its core reasoning engine, while leveraging other models for specialized functions: Gemini for deep research, Nano Banana for image generation, Veo 3.1 for video, Grok for speed-sensitive tasks, and ChatGPT 5.2 for long-context recall. This approach aims to productize and secure the powerful but chaotic workflows pioneered by tools like OpenClaw (formerly ClawdBot), which allowed AI agents to operate locally with extensive system access but was prone to errors and security vulnerabilities. By moving execution to the cloud and using a walled garden of verified integrations, Perplexity Computer seeks to offer the transformative potential of agentic AI—automating complex knowledge work—while mitigating the risks associated with unregulated local agents that could accidentally delete files or execute malicious plugins.

Key Points
  • Uses Claude Opus 4.6 for core reasoning and delegates to specialized models (Gemini, Veo 3.1, etc.) for specific subtasks.
  • Executes in isolated cloud environments with real browser/filesystem access, designed for workflows lasting hours to months.
  • Aims to secure and productize the agentic AI concept, contrasting with the risky, local 'Wild West' of tools like OpenClaw.

Why It Matters

Moves multi-agent AI workflows from experimental setups to a managed product, enabling complex automation while addressing critical security and reliability concerns.