Perplexity announced Wednesday a new product called Teammates that allows multiple AI agents to work together, pick up where earlier sessions stopped, and keep operating without needing fresh prompts. The company says the system can coordinate agents working simultaneously while maintaining shared project state and handling different access permissions. The launch builds on infrastructure Perplexity developed throughout the year, now applied to workflows where agents delegate tasks and continue across sessions.
Perplexity reports that fewer than 2% of submissions get rejected due to being based on stale repository state. The typical submission delay sits under 1.7 seconds, while the 99th percentile remains below eight seconds. The system separates ongoing project state from individual execution environments, letting agents resume work on the same project without relying on a single continuous session. Teammates uses Git-backed repositories to preserve project state across sessions, with each Computer project getting its own repository while agents operate in separate sandboxes and push their changes back to the shared project. When agents work on identical code simultaneously, one agent's commit can force another to reapply its work against the newest version.
"Stateful systems are incredibly hard to build," said Nate Kupp, Perplexity's vice president of infrastructure, pointing to the durability and uptime demands of running millions of sandboxes. The company handles memory separately through its Brain system, which stores what agents have learned in dedicated repositories, allowing background agents to refresh shared knowledge without disrupting changes to project files. Each teammate operates in its own isolated environment and browser session, with access controlled through Computer's existing permissions and credentials secured through encryption, and users can watch progress and step in when necessary.
The technical architecture rests on Firecracker microVMs and Btrfs snapshots, enabling Perplexity to pause, restart, or branch agent sessions without losing saved state. Developers can link tools like Slack and GitHub and configure triggers that signal agents when to take action. The system faces complexity when agents work on the same code at once, since one agent's commit can invalidate another's work. It's unclear how the system stops agents from reaching resources beyond their permissions when tasks get handed off, and for enterprises that uncertainty may shape how much independence they're willing to grant agents without requiring human sign-off. Teammates is initially offered on the web for Perplexity Max subscribers, with iOS support on the roadmap. The question of permissions during task handoffs and what happens when work stalls—particularly when one agent waits on another or acts on outdated information—could determine whether enterprises trust the system with autonomous workflows or insist on tighter human oversight at key decision points.