Perplexity Gave Every Agent Its Own Computer. That Is the Design Decision.

Perplexity launched Teammates on 4 November as a research preview for individual Max subscribers. The product is a persistent agent that owns an ongoing piece of work: it plans tasks, runs them in parallel, checks results, and brings finished output back for review. It sits on top of Perplexity Computer, the multi-agent environment the company launched in February, and adds the layer Computer did not have, which is continuity.
The pitch is easy to summarise and easy to be cynical about. Every agent company now sells "an agent that keeps working while you sleep". What makes this release worth a closer look is one implementation choice, visible in the company's own documentation: each teammate runs on its own virtual machine, with its own desktop, its own browser and its own cookies, separate from other teammates.
Isolation is a security decision, not a feature
Most multi-agent products run several agents in one shared session with different personas and different tool allowlists. It is cheaper and it is where the bugs come from, because a shared browser profile is a shared cookie jar, and a shared cookie jar means one agent's authenticated session belongs to all of them.
Giving each teammate a separate machine changes the failure mode. An agent that wanders onto a page it should not visit takes its own session with it. Credentials are stored encrypted and entered the way a person would enter them, so the agent cannot read them out of its environment. Perplexity inherits tool and Project permissions from Computer, and for GitHub the documentation recommends granting only the repositories the teammate needs.
There is a user-facing side to the same design: you can watch a teammate's desktop, take it over mid-task, or redirect it. That is a supervision interface rather than an autonomy interface, and it tells you what the company thinks the review loop should look like.
How it is packaged
Perplexity's framing is deliberately domestic. A teammate has a name that reflects the project, lives in the left navigation, connects to tools, and takes instructions in a sidebar thread. On day one you describe the work it owns: sources, deliverables, timing, and a schedule if the job repeats. Connectors include Slack, GitHub, Linear, Gmail and Datadog.
The examples in the launch post are specific. A sales teammate notices a regular customer has stopped transacting, reviews their order history and support issues, works out what changed, and either drafts a relevant offer or flags problems to resolve before anyone reaches out. An operations teammate learns that an agency needs two weeks' notice, and builds that lead time into future schedules without being reminded. An engineering teammate finds bugs on a website and hands the findings to another teammate that prepares the fixes.
Adding a teammate to a Slack channel is the part that will decide adoption. Colleagues can assign work or ask for status without opening Perplexity, which is how an agent becomes part of a process instead of a tab somebody remembers to check.
Model choice, and the default
New teammates default to Anthropic's Claude Opus 5.5. You can assign a different model per teammate, send different tasks to different models, and subagents started by a teammate inherit its model unless told otherwise.
That default is the tell. Perplexity has spent two years marketing itself as a router that picks the best model per query. For persistent work, it picked one model as the default and exposed the choice as a per-agent setting. Routing per query optimises for a single answer. Owning a project for three weeks requires a model that holds context, and the company decided the safest default was the most expensive one it could get.
What memory actually holds
Perplexity says a teammate remembers project goals, task history, shared files and feedback across runs, and that each new task builds on the work already done. The examples are mundane and therefore credible: the two-week vendor notice, the tone a team uses in outreach, the list of accounts a sales teammate already researched.
The open question is permissions on memory. Research preview, no documented audit log for what a teammate recalled and why, and no published retention controls beyond what Computer already offers. For an agent reading an inbox and drafting replies, the permission scope granted at connection time is the control that matters, and it is set once, early, by whoever set the teammate up.
Access details worth knowing before planning around it: individual Max subscribers only, at $200 a month or $2,000 a year, web first with iOS promised. Usage draws on Max's monthly credits, and the launch is running without an additional charge during a promotional period, with an undisclosed number of credits required later. Enterprise Pro and Enterprise Max are not included by default, which is an odd gap for a product whose examples are all workplace workflows.
The review burden is the honest test
Every vendor in this category ships a review step. The question that decides whether Teammates is useful is not whether the review step exists but whether reviewing the output is faster than doing the work.
Perplexity's answer is structural. Output arrives as drafts with progress updates and approval requests along the way, and anything consequential, a refund being the example in the documentation, requires a check first. If the drafts are good, a person reviews and approves, and the economics work. If the drafts need rewriting, the person has added a reading task to a writing task, and the agent has made the job slower while appearing to help.
You cannot tell which it is from a launch post. The advertised roles, which per press coverage include Ads Manager, Engineer, QA Engineer and Legal Research Assistant, describe an offering rather than a track record. The only useful trial is narrow: give a teammate one recurring job with a visible output, for two weeks, and count the minutes spent correcting it.
What this release says about the category
Three things are converging. The agent is being packaged as a colleague rather than a tool, because a colleague is something you delegate to and a tool is something you operate. The unit of isolation is being pushed down to the machine, because shared sessions are a security liability that enterprises will eventually refuse. And the price of admission is being set high, at a consumer subscription tier, because persistent agents burn credits continuously rather than per request.
Perplexity spent two years being the company that answers questions. Teammates is the company betting that the question was never the product. Whether that bet pays depends on a number nobody has published yet: how many minutes a week a Max subscriber spends approving work instead of doing it.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.