Perplexity launches Computer agent

Perplexity has launched “Computer,” a new product that turns its search-and-chat interface into an orchestrating agent that can break down a goal into subtasks and assign them to other agents running different AI models. According to Ars Technica, the tool is available to Perplexity Max subscribers and is pitched as something that can “create and execute entire workflows,” potentially running for hours or even months inside a cloud environment with a real browser, filesystem and prebuilt integrations.

The selling point is not a single supermodel but a routing layer: Claude Opus 4.6 for core reasoning, Gemini for deep research, ChatGPT 5.2 for long-context recall and broad search, and other models for images, video and “lightweight” tasks. That architecture also changes where failures land. If an agent chain books ads, emails customers, edits files and makes purchases, the user is no longer clicking the buttons; the vendor is selecting the model, the tools, the environment and the permissions. When the wrong email gets sent, the wrong invoice gets paid, or a prompt-injected web page persuades the agent to leak data, the question is less “why did the user do that” and more “who designed the system that made it easy.”

Perplexity is explicitly positioning Computer as a safer, more controlled alternative to the recent “computer-use” wave—tools that operate browsers or desktops on a user’s behalf. Ars Technica describes OpenClaw, a predecessor in spirit, as powerful but error-prone and exposed to a “Wild West” plugin ecosystem, including at least one reported case where an agent deleted a user’s emails. Perplexity’s answer is a walled garden: the work happens in the company’s cloud, in isolated compute environments, with curated integrations rather than arbitrary plugins. That reduces some classes of risk while expanding others. A centralised agent platform becomes a choke point: compromise the vendor, the integration, or the workflow templates, and many customers inherit the same problem at once.

As agents move from suggestion to execution, the likely response is tighter access control and more aggressive sandboxing. Enterprises that currently tolerate employees pasting data into chatbots will be pushed toward standardised, auditable environments where agents can be constrained by policy—what they can read, what they can write, what they can purchase, and what they can send outside the organisation. The more “Computer” looks like a worker, the more corporate IT will try to treat it like one: least-privilege permissions, segmented networks, approval gates for money movement, and logging that can survive a vendor dispute.

Perplexity’s product is marketed as convenience—one prompt that becomes a campaign, an app, or a research workflow. The operational reality is that the convenience comes from shifting the clicking, the tool selection and the execution environment into someone else’s hands.