Perplexity Computer: Tasks, Connectors, and Safe Ways to Use It

Perplexity Computer is an agent that can research, create files and applications, use connected services, and run recurring tasks. Instead of returning only an answer, it can combine several steps into a workflow and produce a deliverable for review.

Perplexity describes the system as using multiple models, tools, and specialist agents. Exact model assignments and plan access can change, so use the current product interface and official documentation rather than assuming that a named model always handles a particular kind of task.

What Perplexity Computer can do

According to Perplexity's Computer overview, supported workflows include web research, document and presentation creation, application building, email management, and scheduled monitoring. A request may be broken into subtasks and routed to appropriate tools or agents.

Good starting tasks have a clear output and a review point:

• research a market and return a report with source links;
• analyze a supplied spreadsheet and create a workbook with visible formulas;
• turn approved material into a presentation draft;
• summarize messages from an authorized account and draft replies;
• monitor a defined source on a schedule and notify you when a condition is met.

A broad command such as “run my business” leaves too many decisions unstated. Define the objective, approved sources, actions the agent may take, and what must wait for you.

Connectors expand both capability and access

Computer can work with services such as Gmail, Outlook, GitHub, Slack, Notion, and other available connectors. A connector lets the agent reach information or perform actions allowed by its permissions.

Before connecting an account:

1. Confirm which data the connector can read and which actions it can take.
2. Use the narrowest account and permissions that can complete the task.
3. Keep sensitive folders, repositories, and channels outside the agent's scope when they are unnecessary.
4. Review organization policies before connecting work accounts.
5. Remove or revoke access when the workflow is no longer needed.

A connector does not make every source trustworthy. A message, webpage, or document may contain inaccurate or malicious instructions. State which content is evidence and which instructions must come from you.

How scheduled tasks work

Computer can run monitoring, reports, alerts, and other recurring work from its Automations area. Perplexity's scheduled-task guide explains that a background run may start in a fresh isolated session without the conversation's prior context.

Put durable requirements directly in the automation:

Every weekday at 8:00 a.m., review the approved sources listed below for updates about [topic]. Report only items published since the previous run. For each item, include the title, publication date, source URL, and a two-sentence factual summary. Do not send messages or update external records. Notify me only when at least one qualifying item exists.

After creation, verify the cadence, timezone, next run, sources, and notification channel. Review several early runs for duplicates, missing context, and false matches before relying on the automation.

A practical first workflow

1. Define the deliverable. Specify the file or result, audience, deadline, and acceptance criteria.
2. Provide authoritative inputs. Attach approved files and identify which version is current.
3. Set action boundaries. State whether the agent may only draft, may edit a working file, or may act in a connected service.
4. Request a plan. For a complex task, review the proposed steps before execution.
5. Inspect evidence and artifacts. Open the actual spreadsheet, presentation, app, or message draft; do not judge completion from a summary alone.
6. Record corrections. Turn recurring requirements into project instructions or a reusable skill where appropriate.

For persistent source material and instructions, see how to use Perplexity Projects. For research work, use an AI-assisted research plan to connect each claim to appropriate evidence.

Review actions before they affect other people or systems

Check emails, calendar events, tickets, published pages, and code changes before they are sent or deployed. Verify the recipient, destination, permissions, dates, and commitments. Require explicit approval in the workflow for actions with financial, legal, security, or reputational consequences.

For generated documents, confirm citations and calculations. For code or applications, run tests and review access controls. For scheduled monitoring, check that the task stops or alerts you when required information or authorization is missing.

Multi-model orchestration can widen the system's capabilities, but more models do not automatically produce a correct answer. Evaluate the final result against the task's sources and acceptance criteria, and monitor credit usage and duration for workflows that run repeatedly.