perplexity teammates: Agents That Own Ongoing Work

perplexity teammates launched on November 4, and the pitch is a shift from prompting to delegation: instead of tasking an AI session by session, you assign a named agent an ongoing responsibility, and it plans the work, runs tasks in parallel, monitors your connected tools, and comes back when results are ready for review. Perplexity describes teammates as personal agents built on Perplexity Computer, its multi-model work environment, and the launch is a research preview limited to individual Max subscribers on the web, with iOS planned. Source

This post covers what Perplexity's own announcement, changelog, and help-center documentation actually say, what the orchestration model looks like, and the availability details teams should know before planning around it.

What a teammate is

A teammate is a persistent agent with an assigned role. Perplexity's examples: triaging bug reports, tracking metrics, researching a market, or watching competitors for price changes and launches. The agent keeps its role until you change it, schedules its own follow-ups, and retains memory between conversations, so each new task builds on previous work rather than starting cold. Source

Under the hood, teammates build on Perplexity Computer's existing pieces: Connectors, files, code execution, Projects, skills, and memory. A sales teammate, for instance, can draw on a Project's account research, follow a saved skill for outreach, and run a weekly pipeline review automation, tying them together to find stalled conversations and prepare follow-ups for review. Source

The interaction model is review-first. Perplexity says teammates "present results for review," share progress updates and approval requests as they work, and let users inspect, redirect, or take over the desktop mid-task. Whether that review loop is tight enough for your risk tolerance is a per-workflow question, but the design intent is clearly human checkpoints on top of autonomous execution. Source

The launch post sketches example workflows across several business functions, all framed as prompts you might hand a teammate on day one. A support teammate monitors incoming email and prepares responses grounded in help-center documentation. A recruiting teammate follows new candidates end to end, reviewing applications, sending scheduling links, and preparing interview materials. An operations teammate watches for recurring issues or anomalies, and a marketing teammate keeps campaign content moving by drafting next-round materials for approval. Each example pairs proactive monitoring with a human review step before anything external goes out. Source

The orchestration model

The help center is more specific than the launch post about mechanics. A teammate breaks a goal into tasks, runs them in parallel across multiple agents, and adds agents as the work grows. Each teammate gets its own desktop computer and browser, separate from other teammates, and can operate apps and websites directly, including sites without a Connector. Source

Three details stand out for anyone thinking about agent operations:

• Proactive monitoring: teammates watch connected tools such as Slack, GitHub, and Linear, and can start work when something arrives. The documented example is responding to review comments after opening a pull request.

• Model choice: you pick the model each teammate uses and can direct different tasks to different models, consistent with Computer's multi-model positioning.

• Teammate-to-teammate work: one teammate can pass findings to another, and teammates in the same Project share context, files, and instructions while keeping separate roles. They can also be tagged as apps inside Slack, GitHub, and Linear, so coworkers can hand them work or ask for status without touching Perplexity. Source

The changelog's examples show the monitoring loop in concrete terms: check email and calendar every morning and deliver drafts of whatever needs a reply, track named competitors for price changes, launches, and new hires with a weekly summary, or draft replies to every new support email using your policies and order history while checking with you before any refund. Notice the recurring structure in all three: continuous watch, drafted output, explicit checkpoint before anything consequential. Source

The pattern here is a supervisor agent coordinating parallel workers with a shared project space. We covered the open-source version of this pattern in our async subagents post on LangChain's background workers, and the managed-service version in our managed deep agents post. Perplexity's contribution is packaging it as a named coworker with an inbox, rather than a graph you run.

Access controls and credentials

Per the launch post, teammates inherit Computer's tool and Project permissions, run in isolated per-teammate environments with separate browser sessions, and saved credentials are encrypted and entered securely. When connecting GitHub, you choose which repositories a teammate can access. Source

Those are Perplexity's stated controls, and the documentation does not go deeper into audit logging or approval granularity, so treat the details as a question list for evaluation rather than answered concerns. For any agent that reads your inbox and drafts external replies, the permission scope you grant at connection time is the control that matters most, and that is true here by the same logic we laid out in our agent desktop post on computer-use tools.

One related launch-week addition changes the data-handling picture for Mac users specifically: the fully local Computer mode keeps the model, harness, and execution on-device for supported Apple-silicon Macs, so with your permission tasks can work across local files and Mac apps like Notes and Mail without sending personal data to the cloud. That mode belongs to Computer rather than teammates themselves, but it sits in the same changelog and the same evaluation conversation. Source

Availability: read the fine print

The availability story has three parts, and the middle one is easy to miss:

1. Teammates are a research preview for individual Perplexity Max subscribers, on the web at launch, with iOS planned. Free and Pro users would need to upgrade.

2. The help center states explicitly that Enterprise Pro and Enterprise Max are not included by default in this launch. If your organization evaluates on Enterprise plans, that changes the procurement conversation.

3. Two days after launch, the November 6 changelog added adjacent pieces: fully local Computer on Apple-silicon Macs (Pro, Max, and Enterprise, Mac app 26.42.0 or later, 36 GB unified memory recommended), a Voice mode built on OpenAI's gpt-live-1, and Agent Plugins, a vendor-neutral packaging standard for connectors and skills that can be imported from GitHub or a zip file. Source

Setup itself is five documented steps: create the teammate from the sidebar, name it after the project, connect tools (with repo-level GitHub scoping), describe the work it should own including where work comes from and when it should act, then assign a Project and an initial task. Source

Our read

This section is our assessment, separate from the sources. The launch is notable less for any single capability than for the packaging: proactive monitoring plus parallel sub-agents plus a named, persistent identity is close to how teams already describe a junior hire's job, and Perplexity is pricing and gating it like a seat, not a feature.

Two evaluation questions we would ask before delegating real work. First, review depth: the documented model is progress updates plus approval requests, so map which actions in your workflow need approval before the agent acts, and confirm the tool surfaces approvals at those points. Second, failure visibility: an agent that monitors and acts on its own schedule needs a clear record of what it did while you were away, and the current documentation emphasizes control during execution more than audit after it. Neither question is a criticism of the launch; both are standard due diligence for any always-on agent, and the research-preview label is a fair signal that the answers are still settling.

For individual operators, the practical starting point looks like the documented examples: pick one monitoring job with a clear review checkpoint, like morning email triage or competitor tracking, and measure how much of the draft output you accept unchanged over a couple of weeks. That is a cheaper way to learn the tool's failure modes than handing it a workflow where a bad draft leaves the building.