Perplexity launches Computer, wants AI to run tasks for months, not minutes

A multi-model AI orchestrator that coordinates Claude, Gemini, ChatGPT, and Grok

by · TechSpot

Serving tech enthusiasts for over 25 years.
TechSpot means tech analysis and advice you can trust.

First look: Perplexity AI aims to transform how artificial intelligence handles complex, multi-step tasks. Its new platform, called Computer, serves as a coordination layer for multiple large language models, each selected for specific strengths. The system, available to Perplexity Max subscribers, can plan, delegate, and execute tasks over extended periods – ranging from hours to months – without direct supervision.

Rather than relying on a single model, Perplexity AI's Computer system functions as an orchestrator across multiple models. Anthropic's Claude Opus 4.6 serves as the primary reasoning engine, while Gemini handles deep research tasks. Nano Banana generates images, Veo 3.1 produces video, Grok executes lightweight, speed-optimized tasks, and OpenAI's ChatGPT 5.2 manages long-context memory and broad search operations.

Each process runs inside an isolated cloud compute environment with controlled access to a real file system, web browser, and verified tool integrations – a design intended to balance autonomy with security.

The architecture mirrors workflows already adopted by advanced AI users, who often combine specialized models, connect them to local datasets, and use protocols such as MCP (Model Context Protocol) for context sharing.

Computer formalizes this approach as an integrated feature. Users simply describe a desired outcome – for example, a marketing campaign or software prototype – and the system decomposes the request into subtasks, selecting the optimal combination of models for execution.

The concept is part of a broader shift toward agentic AI, an emerging field in which models can operate semi-autonomously over extended periods rather than responding only to instantaneous prompts. Perplexity AI's approach contrasts with single-vendor systems such as Anthropic's internal collaboration product framework, which relies primarily on one company's model ecosystem.

In practice, the design resembles a distributed AI execution layer that dynamically routes tasks to the most suitable engine – functionally closer to an operating system than a conventional chatbot interface.

// Related Stories

The architecture is informed by earlier experiments in autonomous agents. The viral OpenClaw project – previously known as ClawdBot and Moltbot – demonstrated both the potential and risks of locally persistent AI agents.

OpenClaw allowed users to provide long-term context through files such as USER.MD or MEMORY.MD, enabling the system to operate independently to build websites, scan emails, or modify local files. Despite strong technical performance, the open plugin structure created security challenges, including reported cases in which user emails were deleted without authorization.

The design of Perplexity's system appears to be a direct response to that experience. By operating primarily in the cloud and restricting integrations to a curated set of verified tools, Computer aims to preserve functionality while reducing operational risk. In this framing, OpenClaw represented a highly open agentic AI environment, whereas Computer resembles a more controlled platform model, trading maximal flexibility for safety and oversight.

The shift reflects a broader industry competition to define the future of autonomous AI systems. OpenAI has reportedly hired the original creator of OpenClaw, and CEO Sam Altman has suggested that agent-based frameworks will play a significant role in the company's long-term product strategy.

For now, Perplexity's Computer offers a preview of the next generation of AI workflows: multiple specialized models operating in concert, coordinated by a central orchestrator, and capable of executing complex projects over real-time spans rather than in isolated response cycles.