On November 4, 2026, Perplexity announced Teammates, AI agents that take charge of ongoing work. They carry goals, work history, and shared files over into each subsequent run, and keep handling investigation and revisions. As of November 7, it is a web-based research preview for individual Max subscribers and is not included as standard in enterprise contracts. Perplexity already offers scheduled and event-triggered runs, but what Teammates adds is a mechanism that keeps a work role between runs and coordinates the next steps.
An AI with a continuing role vs. the existing Automations
The existing Automations feature runs a process at a specified time or when an event occurs in a connected app. It runs in the cloud, so users don't need to keep a laptop open. Helper agents are also available, so scheduled runs and processing by multiple agents are not new with this release.
Teammates, by contrast, is described by Perplexity as an agent that keeps the same role until you reassign it. It remembers across runs what it is aiming for, what it has tried, and which files it has used, and it adjusts its plan when priorities change. The design has it not only start individual processes but also take the results of those processes and build the next piece of work.
In a sales scenario the company offers, when Teammates detects that a deal has stalled, it looks into past exchanges and support issues, drafts proposals and possible responses, and asks a human for approval. Rather than ending with a notice that something changed, it hands the same agent the investigation and proposals needed after the change. However, this is a description of how it can be used, not a record of results achieved at companies that have adopted it.
Comparing the official descriptions of the two features on five points as of November 7, 2026, gives the following.
| Point of comparison | Automations | Teammates |
|---|---|---|
| Triggering and continuity | Runs at a specified time or on events in connected apps | Holds a continuing role and adjusts work as changes and priorities shift |
| Context carried over | Many background runs use a new agent; some runs carry over the conversation | Remembers goals, history, and shared files across runs |
| Coordinating multiple agents | Helper agents are available | Breaks work down, inspects the results of parallel runs, and decides on adding agents or changing course |
| Execution environment | Runs in the cloud; the user's device doesn't need to be open | Provides each agent with its own dedicated computer |
| After the work | Saves results and notifies you of notable content | Returns deliverables for review and also handles revisions |
The table separates the Teammates feature description and the current Automations specifications into the same categories. Because Automations also has runs that carry over conversations and helper agents, framing this as a shift from an old feature with no memory to a new one with memory would misread the difference. What distinguishes Teammates is that agents with roles and histories coordinate across multiple runs. The comparison does not establish which is better in speed or success rate.
A loop of decomposition, execution, and inspection
Teammates breaks a goal into tasks, picks the parts that can proceed in parallel, and hands them to individual agents. Each agent runs on its own dedicated computer, and after inspecting the results, Teammates decides whether to add agents or change how the work proceeds. The finished product is returned for human review. Perplexity says it can keep monitoring after publication, so work doesn't necessarily end once a deliverable exists.
The "Project" used among agents is a space where the same work's context, files, and instructions are shared. Even when multiple agents operate in separate environments, they can coordinate using the same Project's information. In Slack, there is also a mechanism for collaborating through monitored channels. The envisioned use is that a human's instructions go to one agent, which then splits up the necessary work and carries it forward.
The default model for a new Teammate is Claude Opus 5.5, and it can be changed per agent. Unless specified otherwise, child agents use their parent's model. In other words, the product's continuity is not determined by a single model name. The explanation covers how memory is retained, how agents are coordinated, and even the process of inspecting results and fixing them.
Still, having a process in place is not the same as being able to complete long tasks reliably. From the company's feature description, no specific figures can be found for long-term success rate, maximum parallelism, or the time from monitoring to response. For now, this should be taken as a description of capabilities in a research preview.
A shared Project, separated operating environments
Within the same Project, information is shared, while computers, browsers, and cookies are separated per agent. So the fact that operating environments are separate does not mean files placed in the Project are withheld from other agents. Information to be used jointly and operating environments to be divided by agent are treated as distinct.
Perplexity explains that it encrypts credentials and keeps them invisible to the AI. But not being able to read secrets directly is a separate matter from what can be done on the connected service. Teammates' permissions follow the settings of existing plugins and the Project. Environment separation alone does not make actions on a service you've authorized safe.
What users decide concretely is, for GitHub, limiting access to the repositories that are needed, and for Slack, specifying which channels to monitor and where to post. It is better not to think about handing over the information needed for research separately from handing over the authority to send results outside. The scope of what to delegate, and who reviews the results that come back, should be matched to the connection settings.
When a shared chat is shown to others, it is read-only for them. Sharing a chat does not by itself grant permission for joint editing or free operation. Also, while there is an explanation that deliverables are returned for review, it is not stated that human approval is always required before every action. The scope of work you entrust is determined less by an agent's name than by the information placed in the Project and the permissions of the connected services.
$200 a month and the conditions for always-on use
As of November 7, availability is limited to web use by individual Max subscribers. It is not included as standard in Enterprise Pro or Enterprise Max, so being a product for business use doesn't mean it can be used right away under an enterprise contract. iOS support is planned. The research preview is still in development, and its behavior may change.
Individual Max pricing is $200 per month or $2,000 per year, with annual billing available only on the web. That subscription fee needs to be considered separately from the allotment of credits that run the processing. Under the current credit terms, individual Max gets 10,000 credits per month, with no carryover to the next month. When purchased, the conversion is $1 per 100 credits. These terms may vary by region, promotions, and other factors.
Teammates draws on Max's monthly allotment, and consumption rises when you add parallel agents, handle large files or codebases, or run automatically at high frequency. Paying the monthly fee does not mean any job can run without limit. Auto-refill is disabled by default; if you use it, you must set it explicitly and can decide the refill amount and a monthly cap.
A low balance also affects whether runs continue. Scheduled Automations runs skip that occurrence when the balance is insufficient, and repeated skips lead to a temporary pause. Restoring the balance does not resume them automatically; the user has to restart them with "Enable" or "Run now." This differs from the explanation that ordinary in-progress tasks resume after the balance is restored, so if you entrust scheduled monitoring to it, checking after a pause should be part of your operations.
What to verify on real projects is how many revisions get requested, how long responses take, and how many credits are consumed. If you can confirm that it can take over work within the permitted scope and budget, including the human review burden, you can hand work you had been requesting one task at a time to an agent that carries it on continuously.