Perplexity Teammates Joins the Race to Own Ongoing Work

Perplexity launched Perplexity Teammates on November 5, entering the contest with OpenAI and Anthropic to turn AI assistants into persistent workplace agents. These agents, software that uses tools to pursue assigned goals, are intended to keep handling responsibilities beyond an individual conversation. The company says they build on existing work inside Perplexity, reducing the need to explain projects again whenever another task starts.

The central contest concerns who owns the continuing relationship between workers, their information, and their unfinished assignments. OpenAI’s Dots and Anthropic’s Claude Duet already pursue versions of that relationship. Perplexity’s pitch centers on giving agents access to context users have accumulated in its products, then letting those agents collaborate around ongoing work.

That changes the buying question. A useful answer can justify opening another chat window, but a standing assignment requires confidence in permissions, judgment, and follow-through. For knowledge workers, the attraction is less repetitive briefing. For employers, the concern is whether accumulated context produces dependable execution without creating another stream of work to supervise.

Perplexity Teammates Moves Beyond Individual Tasks

The launch shifts Perplexity’s workplace pitch from completing requests toward maintaining responsibilities over time.

The initial rollout reaches Max subscribers on the web, according to coverage of the persistent work agents. Native desktop and mobile applications are planned, according to chief executive Aravind Srinivas. Those plans should be treated as future distribution, rather than evidence that the launch already covers every work environment.

Teammates builds on Computer, Perplexity’s system for carrying out work through agents and connected tools. The distinction Perplexity emphasizes is continuity: a teammate receives an ongoing job instead of disappearing from the workflow after answering a request. That is a product promise about how work gets managed, rather than a claim that every underlying capability is new.

The company says teammates can draw on Projects, Memory, and Skills already present in its products. Here, memory means retained context, while skills are reusable instructions for performing particular work. Connected files and earlier sessions also form part of the proposed starting point.

For an existing customer, that offers an understandable advantage. Previous research, established preferences, and familiar project material need not become a separate onboarding exercise for every assignment. Whether the system selects the right material, however, remains distinct from whether it can access that material.

A persistent role also changes the shape of an instruction. Asking for a report has a relatively obvious stopping point. Asking an agent to monitor an account introduces continuing decisions about relevance, urgency, and when another action becomes necessary.

That difference explains why the teammate label matters to Perplexity’s positioning. It invites users to describe responsibilities in the language of work: monitor this, maintain that, and tell me when something needs attention. The interface becomes a place to supervise continuing activity, rather than a collection of isolated requests.

The name does not establish human judgment or accountability. A software agent still needs instructions, boundaries, and a way to expose uncertainty. The practical question is whether a continuing role makes those requirements easier to manage than repeatedly starting new conversations.

Availability also needs careful interpretation. A product aimed at workplace activity is not automatically evidence of company-wide adoption. Access through an individual subscription says little about an employer’s approval, procurement process, or readiness to connect sensitive systems.

The immediate change is therefore specific. Perplexity now offers a way to assign persistent agents inside its existing work environment. Claims about replacing departments, transforming organizational productivity, or delivering reliable independent operation require evidence beyond the launch itself.

Existing Work Context Becomes a Competitive Asset

Perplexity’s strongest strategic argument is that useful work already stored in its products gives new agents a starting advantage.

That puts pressure on competing assistant platforms to preserve continuity across research, planning, and execution. A user who has already assembled relevant documents and corrected an assistant’s assumptions faces extra work when moving elsewhere. Persistent assignments make that switching effort more visible.

The company describes teammates as using existing context and connected tools, while retaining ways for people to inspect and redirect their activity. Its public teammate launch details also describe isolated environments and adherence to Computer’s existing tool and Project permissions. These are company descriptions of the product, not independent findings about its reliability.

The strategic importance comes from the relationship between memory and delegation. Context helps an agent understand why a task exists, what previous decisions matter, and which outcome the user prefers. Without that continuity, each assignment begins with another explanation.

However, accumulated information has value only when the system can distinguish current instructions from old material. A project archive contains abandoned options alongside approved decisions. A useful agent must avoid treating every stored sentence as equally authoritative.

That creates pressure on knowledge workers as well as vendors. People adopting persistent agents must decide which records represent their actual working context. Scattered notes and contradictory documents become operational problems when software uses them to choose its next action.

The same issue appears in a team knowledge base. Information needs ownership, boundaries, and enough structure for others to interpret it. An agent adds another reader, with the added consequence that its interpretation can lead directly to tool use.

Consider a hypothetical product team tracking a launch. Its research notes describe customer requests, its planning document defines commitments, and its chat channel contains tentative ideas. An agent preparing a status update needs to know which source governs the release.

That is an illustrative workflow, not a reported Teammates customer result. It shows why broad access alone does not settle the context problem. More available information can improve an assignment, but it can also introduce conflicting signals.

For Perplexity, the competitive opportunity lies in making prior work reusable without demanding constant cleanup from the user. If every recurring assignment needs manual reconstruction, the promised advantage weakens. If the system reliably carries forward relevant decisions, continuity becomes a reason to stay.

The commercial pressure falls most directly on platforms asking users to begin their working context again elsewhere. Feature lists can converge quickly. A history of useful corrections, approved workflows, and project-specific knowledge is harder to reproduce during a casual trial.

That does not establish a lasting advantage for Perplexity. Competing systems can also retain information and connect workplace tools. The launch makes continuity an explicit part of the contest, while leaving its practical quality open to comparison.

The Agent Race Is About Continuing Responsibility

Perplexity is entering a market where persistent work is already becoming a common product promise.

OpenAI introduced Dots in September, describing agents designed to pursue goals over time. Reporting on those always-on agents placed them in an expanding contest over proactive assistance. Perplexity’s launch follows that direction rather than establishing the category by itself.

Anthropic’s Claude Duet adds another reference point. It has been presented as a proactive assistant for professionals and small business owners. The overlap matters because users increasingly encounter similar promises: retain context, monitor activity, and advance work without another prompt.

For buyers, these descriptions can blur meaningful differences. A reminder, a scheduled report, and an agent deciding when to act all operate in the background. They create different expectations about discretion, however, and need different ways to assess success.

A scheduled report follows a defined instruction at an expected time. A standing responsibility can require deciding whether new information changes the assignment. That judgment is where a persistent agent’s usefulness becomes harder to summarize through a feature checklist.

Perplexity’s opportunity is to make its existing workspace the natural home for that responsibility. The relevant competition is the platform a worker trusts to keep an assignment moving. OpenAI and Anthropic serve as direct context for that contest, rather than separate stories about model rankings.

This also explains why a launch comparison should avoid declaring a winner from similar demonstrations. Different users connect different sources, permit different actions, and tolerate different levels of interruption. An agent that suits a researcher may be poorly matched to a finance workflow.

Even successful completion needs interpretation. An agent can produce the requested document while missing a changed assumption. Another can stop to request clarification, appearing slower while preventing an inappropriate action.

Those differences make the unit of comparison larger than a single response. Buyers need to consider the full assignment: setup, execution, correction, and follow-up. A polished result loses value if reviewing its supporting work consumes the time delegation was supposed to save.

The category also has an earlier workplace reference in Claude Cowork. Anthropic brought agent capabilities into desktop work before this latest wave of persistent assistants. The progression shows vendors moving from answering questions toward handling increasingly substantial portions of a workflow.

That history does not prove demand for every new agent interface. It does show why Perplexity faces pressure to turn its research and execution tools into a continuing work product. Remaining an occasional destination would leave the ongoing relationship available to competitors.

The contest therefore centers on repeat use under real conditions. Users must find enough value in keeping responsibilities inside the platform to return, correct it, and trust it again. Announcing another agent establishes participation; continuing usefulness determines whether the relationship lasts.

Collaboration Makes Handoffs Part of the Product

A persistent agent becomes more consequential when its output becomes another participant’s next assignment.

Perplexity says teammates can collaborate with other teammates and communicate with coworkers through Slack. Individual threads provide another way to follow their progress. Those surfaces connect agent activity to places where people already discuss work.

The appeal is easy to recognize. A completed research task often creates a follow-up: update a brief, notify an owner, or request a decision. If each handoff requires copying context into another conversation, some of delegation’s benefit disappears.

Perplexity’s launch demonstration provides a concrete illustration. An advertising teammate identifies favorable campaign performance, prepares expanded activity, and passes a contract extension toward legal review. The demonstration presents a chain of responsibilities, rather than merely a generated marketing document.

It should remain clearly identified as a demonstration. The sequence does not establish independently measured campaign performance or prove that the workflow operates reliably across customer organizations. Its value as evidence is narrower: it shows the kind of coordinated work Perplexity wants users to delegate.

Handoffs create their own information requirements. The receiving participant needs to know what happened, which evidence supports the recommendation, and whether the action is approved. Otherwise, collaboration can spread an uncertain assumption through a longer chain.

Imagine a hypothetical support agent noticing repeated complaints about a feature. It drafts a summary for a product agent, which then proposes a change. If the initial grouping was wrong, the second agent can elaborate an error rather than correct it.

That possibility is an analytical risk, not a documented Teammates failure. It explains why multiple agents should not automatically be treated as independent verification. Their work can depend on the same incomplete evidence or the same mistaken interpretation.

A meaningful handoff therefore needs more than a conversational message. It needs enough supporting context for the next participant to challenge the conclusion. The system should help users distinguish a recorded fact, an inference, and a proposed next step.

Slack participation adds another layer. Coworkers need to understand whether they are giving an instruction, discussing a possibility, or approving an action. Those categories often overlap in ordinary conversation, but an agent needs usable boundaries between them.

The coordination promise is strongest when responsibilities are clear. A teammate can maintain a defined job while passing specific questions to a human owner. The promise weakens when users must reconstruct who requested an action and why another agent continued it.

That makes collaboration a central part of the product’s evaluation. Buyers should examine the quality of handoffs alongside the quality of individual outputs. The important result is whether work advances with an understandable chain of decisions.

Permissions Do Not Resolve Every Judgment Call

An agent can stay within its technical permissions and still make a decision its owner did not intend.

Perplexity’s stated controls include isolated environments, existing permissions, review, redirection, and desktop takeover. Each addresses a real part of supervising software that acts. None, by itself, establishes that the system understands the intended boundary of every assignment.

Isolation separates execution environments. Permissions determine which resources or tools an agent can use. Neither answers whether an allowed action is appropriate at a particular moment.

An agent might have permission to edit a document while misunderstanding which version is approved. It might be allowed to post in a channel while presenting a tentative conclusion too confidently. These are hypothetical examples of judgment risk, rather than claims about observed product behavior.

The distinction matters because persistent operation increases the importance of interpretation. A user cannot assume that every new situation matches the conditions present when an assignment began. Policies change, projects pause, and a once-useful instruction becomes outdated.

Broader reporting on agent access risks emphasizes deliberate limits on the tools and data companies expose. That advice applies to the category generally. It does not provide a product-specific security assessment of Perplexity’s implementation.

There is also a difference between remembering feedback and applying it correctly. A correction might concern one client, one document, or one unusual exception. Extending it to future work can help or harm, depending on whether the original conditions still apply.

For that reason, learning from feedback should be treated as a capability to evaluate. The relevant evidence includes whether users can inspect retained preferences, identify mistaken assumptions, and correct them again. The launch claim alone cannot establish those outcomes.

Review introduces a separate practical constraint. Showing activity is useful, but users still need to find the decisions that deserve attention. A long history of tool actions can leave the reviewer with another interpretation task.

Effective oversight would surface consequential choices, unresolved uncertainty, and departures from the original assignment. That is an evaluation standard, rather than a claim that Teammates already delivers every element. It gives buyers a clearer question than whether the interface looks busy.

Usage also matters for continuing work. Buyers should verify how their actual assignments consume available resources before treating persistent operation as dependable capacity. A role that monitors frequently can behave differently from an occasional document request.

The public launch material does not settle comparative reliability, supervision effort, or sustained customer productivity. Those gaps limit conclusions about organizational readiness. They do not erase the product change, but they make sweeping replacement claims premature.

A reasonable reading is that Perplexity has introduced a more ambitious work interface with an explicit supervision model. Its value will depend on how that model handles ordinary ambiguity. Real workplace work contains enough exceptions that this distinction will remain consequential.

What the Next Phase Needs to Show

The next useful evidence will concern sustained assignments, usable oversight, and the effort required to move work between platforms.

The first signal is repeated customer use on defined responsibilities. A credible account should describe the original assignment, the agent’s permitted actions, and the corrections people supplied. That gives readers enough context to judge whether the result transfers to their own work.

Reports of reduced supervision would strengthen Perplexity’s argument for persistence. Reports of repeated briefing, forgotten boundaries, or extensive rework would weaken it. Neither outcome should be inferred from the number of agents created or the activity visible in a workspace.

The quality of the measurement matters as much as its direction. Completed tasks can vary widely in difficulty and consequence. Comparisons become more useful when they describe what a person previously did and what the agent now handles.

The second signal is how oversight behaves during continuing assignments. Buyers should watch for clearer explanations of permissions, memory corrections, and approval boundaries. The important development is whether people can understand and change the agent’s behavior without rebuilding the assignment.

Evidence that corrections remain appropriately scoped would support the product’s continuity claim. Evidence that users must repeatedly undo old assumptions would weaken it. A persistent assistant gains value from history only when that history remains manageable.

This signal also includes the quality of intervention. Taking over an agent’s environment can resolve an immediate problem. A more durable outcome requires the system to preserve the corrected intent when work resumes.

The third signal is whether competing platforms make existing work easier to bring across. Perplexity’s context advantage depends partly on the effort involved in starting elsewhere. Better transfer of instructions, files, and project history would reduce that advantage.

Conversely, deeper reliance on platform-specific context would make the original workspace more important. Buyers would then need to examine how assignments survive personnel changes, disconnected tools, or a decision to switch services. Continuing usefulness and future flexibility belong in the same evaluation.

Independent comparisons should therefore follow complete workflows rather than isolated prompts. Coverage of agent strengths already emphasizes matching tools to the work. For persistent systems, that matching process must also account for supervision and the quality of ongoing decisions.

For readers, the practical next step is to choose a bounded responsibility and document what successful delegation would look like. Record the sources involved, the decisions requiring approval, and the corrections a person makes during the assignment. Compare the resulting effort with the previous workflow before expanding access. Perplexity Teammates deserves attention because it connects retained context with continuing responsibility, a combination that directly affects everyday information work. The useful question is whether that connection reduces the work required to keep an assignment accurate and moving. Follow the evidence from sustained use, and judge the product by the responsibility it handles dependably rather than the number of roles it can be given.