Perplexity Launches “Computer,” an AI Agent That Manages Other AI Agents

Perplexity has introduced “Computer,” a new AI system designed to coordinate multiple specialized AI agents to complete complex workflows — sometimes running for hours or even months.

The feature is currently available to Perplexity Max subscribers and represents the company’s most ambitious step yet into autonomous AI task execution.

Unlike traditional chatbots, Computer doesn’t just answer questions. It breaks down goals into subtasks, selects the most appropriate AI model for each one, and executes them in a controlled cloud environment.

What Is News

Perplexity launched “Computer,” a multi-agent AI workflow system.

It assigns subtasks to different AI models depending on their strengths.

Core reasoning runs on Anthropic’s Claude Opus 4.6.

Gemini handles deep research, ChatGPT 5.2 manages long-context recall, Grok handles lightweight tasks, and others generate media.

The system operates in isolated cloud environments with curated integrations.

Available now to Perplexity Max subscribers.

How “Computer” Works

Users describe a desired outcome — for example:

“Plan and execute a local digital marketing campaign.”

“Build an Android app for research.”

The system then:

Breaks the request into structured subtasks.

Assigns those tasks to different AI models.

Executes workflows in a sandboxed cloud environment.

Each task runs in an isolated compute environment with access to:

A real filesystem

A real browser

Prebuilt tool integrations

This differs from systems that rely on a single AI model. Instead, Perplexity uses a “best model for the task” strategy.

For example:

Claude Opus 4.6 → reasoning engine

Gemini → research

ChatGPT 5.2 → long-context memory and search

Grok → speed-focused task

Nano Banana → image generation

Veo 3.1 → video production

The architecture reflects a belief that no single model dominates every capability.

What Is Analysis

Perplexity Computer appears to be a more structured, safety-focused evolution of the viral OpenClaw phenomenon.

From OpenClaw to Controlled Environments

OpenClaw (originally ClawdBot) demonstrated what autonomous AI agents could do when given deep access to local systems. It operated directly on users’ machines, managing files, emails, and workflows with minimal supervision.

That freedom came with risks:

File deletion

Prompt injection vulnerabilities

Plugin security issues

Unintended system modifications

Perplexity’s approach is intentionally more conservative:

All operations occur in the cloud.

Integrations are curated.

Tasks run in sandboxed environments.

If OpenClaw represented the “open web” of AI agents, Perplexity Computer resembles a managed app ecosystem.

A New Layer in the AI Stack

The launch signals a shift in AI competition.

Instead of debating which foundation model is best, companies are now competing on orchestration.

The strategic question becomes:

Who controls the coordination layer?

By acting as a model-agnostic conductor, Perplexity positions itself above individual AI providers — potentially insulating itself from model commoditization.

If models become interchangeable engines, the orchestrator gains leverage.

Risks Remain

Despite guardrails, risks persist:

LLMs can still hallucinate.

Multi-agent systems compound errors.

Long-running workflows increase the impact of mistakes.

Data integrity depends on backups and verification.

Additionally, even curated integrations can introduce new attack surfaces if compromised.

Autonomy magnifies consequences.

Industry Implication

Perplexity’s move reflects a broader trend:

AI is shifting from assistant → collaborator → executor.

OpenAI has hinted at similar ambitions, especially after hiring OpenClaw’s developer. Anthropic’s Claude Cowork is also entering this space.

The competition is no longer about chat performance alone. It is about:

Persistent task management

Model orchestration

Secure execution environments

Enterprise-grade reliability

Bottom Line

Perplexity Computer represents a more disciplined version of autonomous AI agents — less experimental than OpenClaw, more structured and enterprise-ready.

It suggests the next frontier in AI is not smarter models alone, but smarter coordination between models

The era of single-chatbot interaction may be giving way to layered AI systems that operate quietly in the background, executing complex workflows with limited supervision.

Whether this becomes transformative productivity infrastructure — or a new class of high-speed mistakes — will depend on how well these orchestration systems handle reliability and control.

One thing is clear: AI agents are moving from conversation to action.