Perplexity launches persistent Teammates for Max subscribers

• Persistent agents can take responsibility for ongoing work, with Perplexity opening web access to Max subscribers.
• Existing Projects, memory, skills and connected tools carry over, according to Perplexity’s launch thread.
• Each agent gets an isolated computer and inherits existing permissions, with human review and desktop takeover available.

Perplexity says every tool call now runs through code mode, the most interesting detail in its harness rewrite. New teammates default to Claude Opus 5.5. Shared teammate chats are read-only for everyone except their owner.

Teammates coordinate Computer sessions and other teammates. Their assigned role persists until it changes, with memory retained between conversations.

Their starting context comes from the existing Computer setup:

• Projects and files.
• Memory and past sessions.
• Skills and connected tools.

Perplexity says teammates also apply feedback to future work, carrying corrections beyond the conversation where they were made.

The Slack connector supports channel monitoring with configurable reply triggers. The setup guide describes configuration around:

• The workspace and channels to monitor.
• Conditions for acting, such as particular tags or questions.
• When and where updates should appear.

Proactivity is configurable during setup; the default favors staying quiet and helping when asked. Teammates also tend to ask for approval, according to a launch-day reply.

Coworkers can assign work and ask for updates in the channels a teammate monitors.

Perplexity describes four areas of work in its harness rebuild:

• Optimizing long-horizon tasks.
• Migrating every tool call to code mode.
• Making the harness plugin-first.
• Evaluating every small change against a battery of RL tasks.

The same engineering post describes cross-device continuation and says the first version of the pplx-decider model was created over a weekend entirely from a phone, through delegated training and evaluation.

For a large goal, a teammate follows five documented steps:

1. Break the goal into tasks and identify parallel work.
2. Start additional agents, each with its own computer.
3. Collect and check results before deciding what comes next.
4. Add or redirect agents as the project changes.
5. Return finished work for review, such as a pull request, report or configuration change.

Teammates can also hand work to other agents. Those in the same Project share context, files and instructions while retaining separate responsibilities.

Users can select models per teammate and direct different tasks to different models. Spawned agents inherit their parent teammate’s model unless told otherwise.

Each teammate has a separate computer and browser environment. The security documentation specifies these boundaries:

• Browser sessions and local cookies stay scoped to that teammate.
• Installed plugins retain their existing permissions.
• Work inside a Project follows that Project’s permissions.
• Saved credentials are encrypted and entered securely; teammates cannot view them.

Users can inspect work, redirect an agent or take over its desktop while it runs.

The research preview requires an individual Max subscription. The preview documentation excludes Pro, Enterprise Pro and Enterprise Max by default.

Public access starts on the web, with native mobile and desktop apps promised later.

Teammates consume Max’s monthly credit allocation. Parallel agents, large codebases and frequent automations can increase consumption; Perplexity’s credit guide values 100 credits at $1 and documents additional purchases and auto-refill.

Asked how many agents to start with, Perplexity’s suggested starting point was two or three.

Perplexity says it already uses teammates for six kinds of recurring work:

• Bug triage.
• QA loops.
• Security monitoring.
• Data analysis.
• Managing production A/B tests.
• Growth marketing.

The company reported about 10,000 teammates in its internal workspace over the preceding 30 days, with roughly 1,000 active the previous day, in its internal usage disclosure. Weekly monitors can fall outside that daily count, as a follow-up explains.

The chart’s methodology defines an active employee as someone who sent at least one prompt to a teammate that week. Agent-hours sum wall-clock runtime across teammate, subagent and automation turns, including parallel runs and tool waits.