Perplexity is moving deeper into agentic AI with Computer, a tool built around a simple premise: instead of asking one model to do everything, give a user goal to a system that can break the work apart and assign pieces to different AI agents.
The product is currently available to Perplexity Max subscribers. Perplexity describes it as “a system that creates and executes entire workflows” and says it is “capable of running for hours or even months.”
What Perplexity Computer is meant to do
Computer starts with an outcome, not just a prompt. A user might ask it to “plan and execute a local digital marketing campaign for my restaurant” or “build me an Android app that helps me do a specific kind of research for my job.” From there, the system works out subtasks and assigns them across multiple agents.
The important shift is orchestration. The user is not directly choosing every model or wiring together every step. Perplexity says Computer coordinates agents running various models and uses the models it considers best suited to each part of the job.
That makes the product different from a standard chatbot session. A chatbot usually responds inside one conversation. Computer is framed as a workflow system that can keep operating across a broader chain of work, with access to tools and a place to run tasks.
A multi-model approach
Perplexity is not relying on a single model provider for the whole system. The core reasoning engine currently runs Anthropic’s Claude Opus 4.6. Other models are used for more specialized work.
According to the source article, the division of labor currently includes:
- Gemini for deep research.
- Nano Banana for image generation.
- Veo 3.1 for video production.
- Grok for lightweight tasks where speed matters.
- ChatGPT 5.2 for “long-context recall and wide search.”
This is the core bet behind Perplexity Computer: different models may be better fits for different tasks, so the workflow should route work accordingly. That contrasts with products like Claude Cowork, which only uses Anthropic’s models.
The model mix also reflects how some advanced users already work. They move between models, choose one for research and another for writing or coding, and connect those systems to data and applications. Computer packages that general pattern into a product instead of leaving users to assemble it themselves.
Why the cloud setup matters
Perplexity Computer runs in the cloud and uses prebuilt integrations. Perplexity says, “Every task runs in an isolated compute environment with access to a real filesystem, a real browser, and real tool integrations.”
That matters because agentic AI becomes more consequential when it can do more than answer questions. A system with a browser, filesystem, and integrations can potentially take action across files, services, and applications. The appeal is obvious: less manual setup, more automation, and a workflow that can carry tasks forward.
But that same capability is also why the design choices matter. A system that touches files and tools needs boundaries. Perplexity’s approach is to put the work inside cloud infrastructure and limit the integrations to a curated environment.
In plain terms, Computer is not trying to be a fully open local agent that can be extended by any plugin a user finds. It is a more managed version of the same broader idea: let AI agents use tools to complete work rather than only generate text.
The OpenClaw shadow
The source article frames OpenClaw as an immediate predecessor to this idea. OpenClaw was previously called ClawdBot and then Moltbot. It was an agentic AI tool that used large language models to operate as a background process on a local machine.
Its range was broad. The article describes it as being able to sort through email history, build websites, and perform many other tasks. With the right permissions and plugins, it could create, change, or delete user files and affect a local system beyond what most users could do with existing models and MCP (Model Context Protocol).
Users gave it longer-running context through files such as USER.MD, MEMORY.MD, SOUL.MD, or HEARTBEAT.MD. Those files helped define goals and guide independent work, sometimes for long stretches without direct input.
That made OpenClaw both compelling and risky. It offered an early glimpse of the kind of knowledge work that AI agents might eventually handle. It was also prone to serious mistakes and exposed to prompt injection and other security problems, partly because of unverified plugins.
The article notes that the same toolkit used to create a viral Reddit clone populated by AI agents was also, in at least one case, responsible for deleting a user’s emails against her will.
Perplexity’s safer lane is still not risk-free
Perplexity Computer is designed to contain some of that risk. Its work happens in the cloud rather than directly on the user’s local machine. It also operates inside a more controlled environment with a curated list of integrations.
The tradeoff is clear. Users may get fewer possibilities than with a wide-open local agent system, but they are also not being asked to trust unverified packages with access to their system. The source article compares the difference, imperfectly, to the open web versus Apple’s App Store.
That does not eliminate the need for caution. Large language models still make mistakes. If Computer is working with data that is not backed up elsewhere, or if users do not verify its outputs, errors could have real consequences.
The broader direction is hard to miss. Perplexity Computer is an attempt to refine the power seen in OpenClaw and package it for a wider audience. It also suggests that the next stage of AI products may be less about one assistant answering one prompt, and more about systems that assign work, select tools, and keep workflows moving over time.
The source article also notes that OpenAI hired OpenClaw’s developer, with CEO Sam Altman suggesting that some of what appeared in OpenClaw will be essential to the company’s product vision moving forward. Perplexity is unlikely to be the only major AI player trying to turn that idea into a more polished product.