Perplexity's Teammates Gives AI Agents Persistent Roles and Memory

Perplexity has launched Teammates, a feature built on its Computer agent platform that lets Max subscribers create persistent, named AI agents for recurring work. Each teammate can retain context, run tasks in the background, coordinate with other agents, and contact people in Slack when a decision requires human approval.

A user assigns each teammate a role, standing instructions, tools, and escalation rules. Perplexity suggests roles such as On-call Engineer, Legal Research Assistant, and Storefront Manager. The structure reduces repeated setup for workflows that span days or continue indefinitely, while human operators retain responsibility for access and decisions.

Persistent roles inherit existing context

Computer is Perplexity’s orchestration platform for complex workflows. The company says it breaks projects into smaller tasks, selects among 19 AI models for each task, and delegates work to specialized agents. Teammates adds persistent memory, proactive monitoring, and communication between agents to that system.

According to Perplexity, each teammate inherits the user’s existing Computer projects, skills, plugins, memory, and permissions. An agent can begin with established coding conventions, brand guidelines, or contract terms already in the workspace. That inheritance can shorten setup, but it also makes permission review essential before an agent starts operating.

Five mechanics shape the product

• Isolated environments. Each teammate receives a separate desktop, browser session, and local cookie store. A marketing agent can remain signed in to Meta while an engineering agent works in GitHub, with each session kept separate.
• Persistent threads. Every teammate retains a continuing thread in the sidebar. Users can inspect its activity, decisions, and prior context across multiple runs.
• Proactive monitoring. Teammates can watch assigned systems and respond when conditions change. In Perplexity’s demo, an On-call Engineer detects a spike in HTTP 502 errors after a deployment, traces the problem to a dropped retry budget, and opens a pull request with a rollback plan based on earlier merged work.
• Agent handoffs. Teammates can send work to one another. Another company demo shows an engineering agent fixing a password-reset crash, passing the pull request to a QA agent for desktop and mobile testing, and receiving the test results for review.
• Slack escalation. Teammates can contact coworkers in Slack when they need approval, clarification, or another decision, allowing routine work to continue outside an active chat session.

Humans can inspect, redirect, or take over

Autonomous browser and desktop access creates operational and security risks, so Teammates uses the permissions already configured for Computer tools and projects. Users can inspect work in the persistent thread, redirect an agent during a task, or take control of its desktop.

Escalation rules determine when a teammate should involve a person. Perplexity’s on-call example pages a human when a proposed fix needs approval or when the agent cannot isolate a cause within 15 minutes. Teams will need to define those thresholds carefully because inherited permissions and broad escalation rules can expand an agent’s practical reach.

Price, platforms, and credits

Perplexity has not specified the size or expiration date of the promotional grant. Credit consumption for sustained monitoring, browser use, and multi-agent workflows will determine the practical cost once that grant ends.

Durable roles change the agent workflow

Perplexity’s architecture treats model selection as an orchestration problem. Computer routes subtasks among specialized models, while Teammates packages those capabilities into durable roles with memory, tools, permissions, and responsibilities.

Developers adopting the product would manage recurring workflows as persistent agents with credentials, retained state, escalation policies, and inter-agent messages. That model raises familiar production concerns around secrets management, audit logs, failure recovery, approval gates, and budget controls.

Small teams could assign parallel roles to sales follow-up, deadline tracking, competitor monitoring, incident response, and invoice reminders. Standing briefs would drive routine work, while defined checkpoints would route sensitive actions to a person.

Production use still has tests to pass

• Long-running reliability: Whether agents retain accurate context, recover from failed tools, and avoid compounding earlier mistakes.
• Handoff integrity: How downstream agents respond when another teammate supplies incomplete or incorrect work.
• Permission boundaries: Whether teams can limit access to credentials, repositories, customer data, and production systems with sufficient precision.
• Auditability: Whether persistent threads provide enough detail to reconstruct actions and decisions during an incident or review.
• Cost predictability: How quickly monitoring, browser sessions, retries, and agent-to-agent workflows consume credits.
• Alert quality: Whether Slack escalations reach people at useful moments without producing excessive notifications.

Teams evaluating Teammates can contain early risk by starting with bounded workflows, granting the least access required, requiring approval for consequential actions, and measuring task success, latency, intervention rate, and credit use. Production results will show whether persistent roles deliver dependable automation beyond Perplexity’s launch demonstrations.