Perplexity puts Teammates in charge of 5 workflows without a new prompt

Most AI agents today resemble a highly capable intern on their first day at work. You can assign them a task, but first you must explain the context again, open the necessary tools, and list every subsequent step. Perplexity is trying to eliminate precisely this routine: its new Teammates receive not a standalone request but an ongoing area of responsibility.

An agent can remember project goals, the history of completed tasks, and human feedback. It monitors changes, launches new actions without a fresh prompt, and returns the result for review. The company highlights five main use cases: marketing, sales, support, operations, and finance.

What this looks like in a routine task

Suppose a regular customer starts ordering less frequently from an online store. A conventional chat will notice nothing until a manager gathers the data and asks a question. In Perplexity's scenario, a Teammate tracks the change, examines order and support-contact history, identifies a possible cause, and prepares an appropriate offer. If an old complaint must be resolved first, the agent tells a human and waits for approval before taking further action.

This is more than a long conversation with a model. A Teammate coordinates Perplexity Computer sessions, uses project files and tools, and delegates part of the work to another agent when necessary. For example, one agent finds errors on a website while another prepares fixes. A Slack connection lets colleagues assign tasks and ask about progress directly in a shared channel.

During setup, the user specifies where work comes from, what outcome is required, and when the agent should act. They then connect plugins such as Gmail, Slack, or GitHub and select the repositories the agent may access. Each Teammate operates in a separate environment and receives its own browser sessions. Perplexity says stored credentials are encrypted; a human can inspect the work, redirect the agent, or take control of the desktop.

Where the real value lies

The main shift is from automating a single command to overseeing a process. A support agent can review complaints daily, identify recurring issues, draft replies, and track whether promised fixes have been completed. An operations agent can check deadlines and send reminders about missing documents 72 hours before the deadline. A marketing agent can monitor campaigns and prepare new ad variants based on the results.

For a product team, this means an agent can be assigned an entire cycle: find a problem in the sign-up funnel, collect data, propose a change, submit it for review, and later measure the effect. But such a scenario is useful only with clear boundaries. The agent must know which actions are permitted automatically, where confirmation is required, and what signal should stop the work.

What remains unproven

The announcement describes the product's capabilities but includes no independent tests, accuracy metrics, or data on time savings. It is unclear how reliably Teammates retain context in long projects, how often they choose the wrong next action, and how much it costs to keep several agents running continuously. The Hacker News discussion was also skeptical: users recalled errors in earlier Perplexity products and questioned how the offering differs from competitors' solutions.

There is also a risk common to agentic systems: the greater the access, the more costly the mistake. An agent with access to email, a repository, and working files can not only accelerate a process but also send an inappropriate message, modify the wrong object, or expose unnecessary data. Separate environments help, but they do not replace least-privilege access, an audit log, and mandatory confirmation for irreversible operations.

Teammates are already available in the web version to Perplexity Max subscribers; the company says iOS support will be added later. The practical approach is straightforward: choose one recurring process with a measurable outcome, give the agent only the permissions it needs, and run a pilot. The comparison should focus not on how impressive the demo looks, but on the share of tasks completed correctly, the number of human interventions, the cost per outcome, and the consequences of the worst plausible error.