Perplexity dropped Teammates on Wednesday, a new orchestration layer that lets multiple AI agents coordinate tasks, remember session history, and operate asynchronously without constant human prompting. The system builds on the company’s existing Perplexity Computer infrastructure, separating persistent project state from individual execution environments to allow agents to resume work exactly where they left off.

The Stateful Sandbox Problem

Nate Kupp, Perplexity’s VP of Infrastructure, told The New Stack that maintaining reliable systems at this scale is brutal. "Stateful systems are incredibly hard to build," Kupp said, referencing the durability requirements of running millions of sandboxes. To solve this, Teammates relies on Git-backed repositories for shared workspaces. Each project has its own repo, while agents operate in isolated sandboxes and submit changes back to the shared state. If two agents edit the same code simultaneously, the system forces reapplication of changes against the latest version, though Perplexity claims fewer than 2% of submissions are rejected due to outdated state.

MicroVMs and Memory Isolation

The underlying tech, dubbed SPACE, uses Firecracker microVMs and Btrfs snapshots to enable pausing, resuming, and forking agent sessions without losing state. This architecture allows for sub-1.7-second median submission latency, with p99 staying under eight seconds. Memory is decoupled from file changes via Perplexity’s Brain system, which stores learned knowledge in dedicated repositories. This separation ensures that background agents can update shared knowledge bases without corrupting the project’s active file tree.

Permission Ambiguity in Multi-Agent Flows

While Teammates supports integrations with Slack and GitHub for triggering agent actions, the permission model raises red flags. Each teammate runs in an isolated browser session with encrypted credentials, but it remains unclear how the system prevents an agent from accessing resources beyond its defined permissions during task handoffs. For enterprise users, this ambiguity could limit how much autonomy they are willing to grant agents without requiring human approval checkpoints.

Key Takeaways

  • Teammates is initially available only on the web for Perplexity Max subscribers, with iOS support planned for later.
  • The system uses Git for state management and Firecracker microVMs for session persistence, achieving under 2% rejection rates for outdated submissions.
  • Permission inheritance during agent-to-agent handoffs remains undocumented, creating potential security risks for enterprise deployments.

The Bottom Line

Perplexity is solving the hard engineering problem of stateful agents, but until they clarify how permissions propagate across handoffs, Teammates feels like a power tool missing its safety guard. The focus on infrastructure durability is impressive, but the lack of transparency on access control during multi-agent coordination is a major hurdle for widespread enterprise adoption.