Perplexity Teammates is a new attempt to turn an AI assistant into a colleague who has an ongoing job rather than a blank chat box waiting for instructions. Announced in early November 2026, the feature lets eligible users assign an agent a continuing task, give it access to selected tools and review what it has done. The appeal is obvious to someone who has spent a morning chasing emails, checking figures and reminding colleagues about unfinished work. Whether the system can perform those duties reliably is a separate question.
In its launch explanation, Perplexity describes agents that coordinate work over time, remember project context and follow up as circumstances change. A sales agent might notice that a customer has stopped ordering, examine the account history and prepare a response for review. An operations agent could watch deadlines and documents. The examples show the intended workflow, but they are not an independent measurement of how often the system reaches a correct conclusion.
What Perplexity Teammates do between prompts
The basic distinction is persistence. A conventional assistant can write a draft when asked. A teammate is assigned an area of responsibility and may continue checking conditions or moving a project forward without a fresh instruction for each step. Perplexity says a teammate can coordinate Computer tasks, use project files and tools, pass work to another teammate and share progress updates. That sounds closer to a project coordinator than a search engine, although the finished work still requires people to judge whether it is useful.
The most promising examples are routine activities where the next action is relatively clear. Checking for missing documents, collecting draft responses and preparing a weekly summary do not necessarily require a person to make every small intermediate decision. On the other hand, a business process full of exceptions may produce awkward choices. An assistant may identify a customer whose order frequency has dropped without understanding a previous complaint or an agreement made on the telephone. ‘Still working’ is worthwhile only if it is working on the right thing.
An assistant with access can also make a bigger mistake
Persistent agents may need access to email, calendars, documents and other business applications. Each connection makes the assistant more useful and expands what it can see or potentially change. That is why permissions, approvals and audit trails matter as much as the model’s ability to write a persuasive message. Our earlier explanation of human oversight for AI agents applies to a teammate that may be running when its owner is busy with something else.
Perplexity says teammates use existing Computer permissions and run in separate environments. A 6 November product changelog specifies that the first web availability is for Max subscribers. That is a launch condition, not proof that all possible integrations or workflows have been thoroughly tested in every setting. A company considering an agent should start with a bounded task, limited access and a clear expectation about which actions need explicit approval.
The difficult part is knowing when to stop
A useful teammate should recognise the difference between preparing a reply and sending it. Suppose a customer complains about a refund: collecting the previous correspondence and suggesting wording could be helpful, while issuing a refund or making a legal promise would require stronger authorisation. Similarly, spotting a missed deadline is straightforward compared with changing a supplier arrangement to meet it. The stronger the software becomes at taking actions, the more important it is to specify what it must never do on its own.
It also needs a way to recover from stale information. A project can change direction after a conversation in a corridor, a phone call or a document that has not yet been uploaded. The system may continue diligently with the wrong assumption unless the owner can inspect its working context. Existing experiments with AI teammates in meeting workflows illustrate why catching every instruction is not the same as understanding every decision a team has made.
Could this make small businesses more effective?
For a business without a large administrative team, keeping track of recurring tasks has a real cost. A tool that spots missing invoices, flags overdue work and prepares the next sensible message might free people to concentrate on customers. The opportunity is particularly interesting because the agent does not need to be dazzlingly creative to be useful. It must be consistent, accurate enough to save review time and transparent about work it has not completed.
But there is a less attractive scenario in which employees spend their day checking a stream of almost-correct outputs. A system that drafts too many unnecessary emails, repeats old assumptions or generates awkward customer interactions could increase work rather than reduce it. Perplexity has not established that its promotional examples will translate into guaranteed productivity savings for every subscriber. The sensible test is the number of real tasks completed correctly after accounting for supervision and correction.
Who can try it now?
Perplexity says Teammates are initially available to Max subscribers on the web, with wider access and capabilities planned later. Users should check the actual controls in their account and the permissions attached to connected applications before assigning anything consequential. A small trial using non-sensitive records and a task that has an obvious correct outcome is a more informative test than giving an agent broad access on the strength of a demonstration.
The potential shift is from AI that responds to questions towards AI that holds a limited responsibility. If that succeeds, the software may become less visible, because the useful work is done while people concentrate elsewhere. The crucial requirement is that it stays accountable. The best artificial teammate should make it easier to know what happened, what is still pending and who must make the final decision, not create another colleague whose actions nobody quite understands.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.