Perplexity brings AI agents feature called Teammates to the Mac

Perplexity’s deployment of the Teammates agentic framework to macOS represents a significant architectural shift from reactive query-response paradigms to persistent, goal-oriented autonomous agents operating within the Apple silicon runtime environment. The macOS implementation leverages the Neural Engine via Core ML for on-device inference orchestration while maintaining a hybrid execution model that offloads complex reasoning chains to Perplexity’s distributed inference fabric. This architecture utilizes Apple’s App Sandbox entitlements for controlled filesystem and network access, enabling Teammates to execute multi-step workflows — such as codebase analysis, document synthesis, and API interaction — without violating the platform’s security model. The engineering challenge centered on reconciling the agent’s long-running background execution requirements with macOS’s aggressive process lifecycle management, necessitating a launchd-backed persistence layer with explicit user consent flows for background task continuation.

The cross-platform parity achievement between iOS, web, and now macOS required a unified agent state synchronization protocol built atop WebSocket connections with automatic reconnection semantics and conflict-free replicated data types (CRDTs) for conversation history merging. On macOS specifically, the client integrates with the Unified Logging subsystem for telemetry emission and adopts the new Background Assets framework introduced in macOS 15 for pre-fetching context artifacts during system idle windows. Memory pressure handling follows the Jetsam priority band model, with the agent process tagged as a user-interactive daemon to prevent premature termination during active collaboration sessions. The implementation also exposes a local Model Context Protocol (MCP) server endpoint on localhost, allowing third-party developer tools to invoke Teammates capabilities programmatically without cloud round-trips.

Actionable Diagnostic & Implementation Steps

1. Verify the Perplexity macOS application bundle identifier (com.perplexity.mac) holds the com.apple.developer.background-task entitlement via codesign -d --entitlements :- /Applications/Perplexity.app; absence indicates an unsigned or App Store-restricted build lacking daemon persistence.
2. Enable local MCP endpoint exposure by executing defaults write com.perplexity.mac EnableLocalMCP -bool true followed by application restart; confirm listener activation with lsof -i :8080 | grep LISTEN to validate stdio transport readiness for IDE extensions.
3. Configure Jetsam priority band override for sustained agent operation: create ~/Library/LaunchAgents/com.perplexity.teammates.plist with PriorityBand key set to UserInteractive and load via launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.perplexity.teammates.plist.

Ecosystem Compatibility & Support Notice

• Requires macOS 15.0 (Sequoia) or later for Background Assets framework and Neural Engine Core ML delegation APIs.
• Apple Silicon (M1/M2/M3/M4 series) mandatory; Intel Macs fall back to cloud-only inference with no local acceleration.
• Minimum 16 GB unified memory recommended for on-device embedding cache; 8 GB systems experience aggressive memory compression impacting agent responsiveness.
• Network extension entitlement required for WebSocket persistence across network interface changes (Wi-Fi to Ethernet handoff).
• Sandboxed filesystem access limited to user-selected directories via Security-Scoped Bookmarks; full disk access not supported.

Source and technical reference: 9to5Mac.