Perplexity Teammates: Persistent Work Agents on Computer for Max Subscribers

Perplexity has moved from agents you prompt to agents you assign. On November 4, 2026, it launched Teammates, "personal agents that can take on" long-running work: they "plan and coordinate projects, switch gears as priorities change, and present results for review" (Perplexity). They build on Computer, Perplexity's multi-model agent, and start on the web for Max subscribers.

This post uses Perplexity's announcement for facts and a secondary write-up by AlphaSignal for details Perplexity's post does not give. Where the two differ, we say so.

TL;DR: the questions people ask first

What Perplexity says Teammates do

Perplexity describes a teammate as something that "coordinates Computer sessions and other teammates to keep projects moving." Teammates remember project goals, task history and shared files across runs, so "each new task a teammate takes on builds on the work already done."

Three behaviors define the product:

1. Ongoing ownership. You delegate an area of work, such as marketing or support, instead of a single prompt.
2. Proactive monitoring. Teammates "investigate and act without waiting for a fresh query." Perplexity's example: a sales teammate notices a regular customer has stopped transacting, reviews order history and support issues, then drafts an offer or flags problems to fix before outreach.
3. Memory and feedback. If you tell an operations teammate an agency needs two weeks' notice, it factors that into future schedules.

Perplexity's examples by function

• Marketing: manages social posts, email campaigns, ads, SEO and review notes, monitors performance and competitor promotions, and drafts fixes when results slip.
• Sales: finds opportunities in inquiries and unfinished conversations, prepares outreach and quotes, follows up on replies.
• Customer support: reviews the previous day's complaints, summarizes trends, flags cases needing a personal response, and tracks each case to resolution.
• Operations: keeps deadlines, checklists, vendors and paperwork on track, following up on missing items 72 hours before a deadline.

These are Perplexity's own illustrations, not measured results. The post gives no data on accuracy, time saved or error rates.

Slack, handoffs and isolation

You can add a teammate to a Slack channel so colleagues can hand it work or ask for updates. Teammates can also pass work to each other: "One teammate finds bugs on a website and passes its findings along to another to prepare fixes."

On safety, Perplexity says teammates build on Computer's existing tool and Project permissions, each runs in a separate environment with its own browser sessions, saved credentials are encrypted and entered securely, and users can review completed work, take over the desktop or redirect a teammate mid-task.

AlphaSignal adds demo details we cannot confirm from Perplexity's text: an on-call engineer teammate that traces a spike of HTTP 502 errors and opens a rollback pull request, and an escalation to a human if it cannot isolate a cause within 15 minutes. Treat those as reported demo scenarios.

Setup, per Perplexity

1. Sign in on the web and choose Teammates in the left navigation.
2. Name the teammate for the project it will do, then follow setup.
3. Connect tools. Plugins include Slack, GitHub and Gmail. For GitHub, pick which repositories it can access.
4. Describe the work: where it comes from, what it delivers and when it should act.
5. Assign it to a project and give it a first task, such as researching a market or investigating bugs.

Price and credits: what is unclear

Perplexity's post says Teammates are for Max subscribers and that Free and Pro users can upgrade. It does not give a Teammates price. AlphaSignal reports Max at $200 per month with 10,000 monthly credits and a launch promotion in which Teammates usage is free before standard credits apply, and notes Perplexity has not specified the size or end date of that grant. We could not confirm those figures on Perplexity's blog, so check your account.

The practical cost question is credit burn. A teammate that monitors continuously, opens browser sessions and talks to other agents could use credits faster than a one-off Computer task. Measure it before you give a teammate a standing job.

How it compares

Persistent agents are the current frontier for assistants. Our coverage of related launches includes OpenAI's always-on Dots agents, ChatGPT work with a cloud browser, Perplexity Computer running fully local on a Mac and Cloudflare's controls for personal agents. The shared idea is that an agent keeps state, acts on a schedule and reports back, instead of waiting for a chat turn.

What differs is where the agent lives and who bears the cost of mistakes. Teammates run inside Perplexity's cloud environment, share your Computer permissions and escalate through Slack.

Where Teammates fit in Perplexity's agent stack

Perplexity's own blog shows the pieces Teammates stand on, published over the last several weeks:

Read in order, the posts show a deliberate path: sandboxes, shared project state, injection defenses, then agents that own long-running jobs. That context supports Perplexity's claim that Teammates build on Computer's Projects, skills and memory rather than starting from scratch. It does not prove the safety of the whole system. The SPACE and BrowseSafe++ posts describe Perplexity's own tools, and independent testing of Teammates will have to show how well they hold up in practice.

What would change our view

A few facts would make this a stronger or weaker product, and none are in the launch post:

• A clear credit model. If monitoring and handoffs are metered predictably, teams can budget. If not, a standing teammate becomes a surprise bill.
• Granular permissions. Per-repository GitHub access is a good start. Per-action limits, such as "can open a pull request but not merge," would matter more.
• Independent failure reports. Persistent agents tend to fail in slow, quiet ways: stale memory, a bad handoff, a monitoring loop that never ends. Early user reports will show which of these appear.
• Enterprise controls. The launch targets Max subscribers. Admin controls, audit export and policy settings would decide whether teams can adopt it at work.

Until then, treat Teammates as a promising pilot tool, not a replacement for a person who is accountable for the work.

Risks to test before you trust one

• Inherited permissions. A teammate uses the tools and Project permissions already in Computer. Review them before you assign work, and remove access the job does not need.
• GitHub scope. Choose specific repositories. Do not grant all repos by default.
• Escalation design. A teammate is only as safe as its rules for when to ask. Too few approvals let mistakes through; too many create alert fatigue.
• Handoff errors. One teammate can pass flawed findings to another. Check the first several handoffs by hand.
• Prompt injection. An agent that reads email, web pages and tickets can be steered by hostile text. Perplexity published a prompt injection benchmark for browser agents on November 2, called BrowseSafe++, which is a useful sign it takes this seriously, though we have not tested it. For background, see our MCP security guide.
• Audit trail. Each teammate keeps a thread you can inspect. Confirm it is detailed enough to reconstruct an incident.

For teams that want a runtime check on what agents can do, AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.

A safe pilot plan

1. Pick one low-risk recurring task, such as a weekly summary of support complaints.
2. Grant read-only access where possible and one Slack channel.
3. Require approval for anything that sends external messages or spends money.
4. Track success rate, how often you intervene, and credits used for two weeks.
5. Expand only if intervention rate falls and the audit trail is clear.

Teammate brief template:
Role: <what you own>
Inputs: <where work comes from>
Allowed tools: <list, with repo or channel limits>
Deliverable: <format and cadence>
Ask a human before: <sending, spending, deleting, deploying>
Stop and escalate if: <time limit or unknown cause>

What this means for what you build or pay

If you already pay for Max, Teammates add a way to run recurring work without re-prompting. If you do not, the plan jump is the cost, plus whatever credits the agents use. If you build agents yourself, the design is worth studying: persistent roles, per-agent isolation, inherited permissions and explicit escalation. It is a reasonable template, and it shows which controls buyers will ask for.

Related reading

• OpenAI Dots: always-on agents
• ChatGPT work with a cloud browser
• Perplexity Computer fully local on Mac
• Cloudflare personal agents controls
• MCP security guide

Details are accurate as of November 5, 2026. Plans, credits and availability may change; check Perplexity's site.