Perplexity Enters Work AI Agent Space: Teammates Goes Always-On

Perplexity has officially joined the race for workplace AI agents, launching an always-on agent product called Teammates that directly targets enterprise offerings from OpenAI and Anthropic. Unlike most assistants on the market that require manual triggers, Teammates emphasizes being "always online" and aims to seamlessly migrate users' existing projects, skills, and memory into agent workflows.

A Positioning Shift from Search Tool to Work Agent

Perplexity's core identity has been AI search, and the launch of Teammates marks its expansion into a productivity platform. According to the original report, the product is explicitly positioned as "an AI agent built for work," going head-to-head with OpenAI and Anthropic on the same track. This means Perplexity is no longer content to be an information gateway—it wants to become the task execution layer, where users set goals and the agent operates continuously in the background.

Always-on operation is the key differentiator. Most workplace AI assistants rely on single-conversation triggers and stop once a task is complete; Teammates is designed to reside persistently in the workspace and handle continuous tasks that span time periods and sessions. This model places higher demands on underlying orchestration and context management, which explains why Perplexity bundles it with Perplexity Computer.

How Teammates Fits into Existing Workflows

Multi-Role Division of Labor and Skill Inheritance

Teammates can take on multiple roles within the same workspace. The original article notes that it can "pull existing projects, skills, and memory from the Perplexity tools you already use"—a point worth unpacking: the agent's capabilities aren't configured from scratch but inherit project context and operational preferences accumulated through the user's historical usage. For teams, this lowers the migration cost of onboarding the agent; for individual users, the agent feels more like an extension of existing work habits than a brand-new tool starting from zero.

Mechanisms for Agent Memory and Context Inheritance

An agent's "memory" is typically divided into short-term conversation context and long-term persistent memory. Long-term memory stores users' historical projects, preferences, and operation records via vector databases or structured profiles, and the agent retrieves this information in new tasks to maintain continuity. Skill inheritance means the agent reuses already-configured toolchains, prompt templates, or workflows rather than rebuilding from scratch each time. This design can significantly reduce migration costs, but memory accuracy depends on retrieval quality and expiration policies: outdated or incorrect memories get repeatedly carried into new tasks, causing bias accumulation. When multiple roles share the same memory pool, namespace or role isolation mechanisms are also needed to prevent context cross-contamination—this is precisely the key design point for product experience.

Perplexity Computer as the Execution Foundation

Teammates integrates with Perplexity Computer, described as "an AI system that can deploy agents on your device." Deploying agents to local devices means task execution can reach local files and environments, not just cloud conversations. This architectural choice sets it apart from purely cloud-based agents and also raises new questions around permissions and security—once an agent resides persistently on a device, defining its accessible scope is a question the product must answer at launch.

Architectural Differences Between Locally Deployed and Cloud-Based Agents

Cloud agents typically run on the provider's servers, interacting with user data via APIs, with computing resources and environments managed uniformly by the provider; local device deployment places the agent runtime on the user's own hardware, allowing the agent to directly read local files, invoke local applications, access intranet resources, with lower response latency and without requiring all data to be uploaded. The trade-off is that the user's device bears the computing load, and the agent's permission scope is tightly coupled with device security policies. Perplexity Computer is the vehicle for this approach, letting Teammates reach local environments that cloud assistants cannot access, but also shifting part of the responsibility for permission isolation, sandboxing mechanisms, and log auditing to users or enterprise IT departments.

Competitive Landscape: Three Distinct Paths

Perplexity's entry has shaped three typical paths in the workplace agent race. OpenAI and Anthropic leverage general-purpose large model capabilities to expand into the agent layer, with their advantage in the models themselves; Perplexity started with search and information processing, with strengths in knowledge acquisition and integration, so Teammates' differentiation will likely center on "information-intensive knowledge work."

From a product strategy perspective, Perplexity has chosen to play the "always-on + multi-role + memory inheritance" combination, avoiding a head-on battle with giants over model parameters and instead innovating on usage patterns. Whether this approach succeeds depends on whether users are willing to let agents reside long-term in their workspaces and share memory—this is both a source of efficiency and a trust threshold.

Competitive Dynamics in the Enterprise AI Agent Market

Since 2024, OpenAI has launched Operator, Anthropic has introduced Computer Use and Claude Enterprise, and Microsoft Copilot and Google Gemini have also accelerated their penetration into the agent layer. Workplace agents have become the second battleground for large model vendors. The differentiation paths roughly fall into three categories: model capability-driven (relying on stronger reasoning and tool calling), ecosystem integration-driven (bundling with office suites and cloud services), and usage pattern-driven (always-on, localized, memory inheritance). Perplexity belongs to the third category, building on its search-origin information integration capabilities while avoiding the model parameter race. Victory in this market often depends on trust and compliance assessments in enterprise procurement rather than pure benchmark scores.

Real-World Challenges for Deployment

The first hurdle for always-on agents is permission boundaries. To pull projects, skills, and memory, the agent must obtain deep account access; once deployed to devices, whether local environment isolation mechanisms are sufficiently clear will directly affect enterprise customers' willingness to adopt.

The second hurdle is task controllability. When an always-on agent runs continuously in the background, how users perceive its status, intervene in its decisions, and revoke its permissions—the original article provides no details. For workplace scenarios, auditability and interruptibility are often more important than capability ceilings.

The third hurdle is memory accuracy. Inheriting historical memory improves continuity, but incorrect or outdated context also gets carried into subsequent tasks, causing bias accumulation. When multiple roles share the same workspace memory, how to design context isolation between roles is a critical dividing line for product experience.

Perplexity has not yet disclosed specific pricing or availability for Teammates, but the direction is clear: pushing AI from "you ask, it answers" to "it stands by on duty." The decisive factor in this competition may not be how powerful the model is, but whether the agent can be confidently delegated.

One-line Summary

Perplexity launches always-on work agent Teammates, competing with OpenAI and Anthropic via persistence, memory inheritance, and local deployment.

Paragraph Summary

Perplexity has released Teammates, an AI agent for workplace scenarios positioned as an "always online" agent product, directly targeting enterprise offerings from OpenAI and Anthropic. Teammates can take on multiple roles within the same workspace and inherit users' existing projects, skills, and memory to lower migration costs; it integrates with Perplexity Computer to support local agent deployment on user devices, distinguishing it from purely cloud-based solutions. The competitive focus shifts from model capability to usage patterns, with always-on operation, memory inheritance, and local deployment forming its differentiated combination. The article notes that permission boundaries, task controllability, and memory accuracy are the main barriers to deploying always-on agents, and Perplexity has not yet disclosed pricing or availability.