agent-observability

Real-time Claude Code agent observability is achieved by capturing deterministic hook events and streaming them in-memory to a WebSocket-based frontend dashboard for live visualization. The dashboard provides swim lanes, session tracking, and live charts to monitor parallel agent execution.

Debugging parallel Claude Code agents is best handled by using a real-time observability dashboard with swim lanes to display concurrent session flows. You can filter events by specific agent, session, or event type to quickly identify bottlenecks or failing tools across multi-agent workflows.

Claude Code observability does not require a database; this approach uses a zero-database architecture that stores events as human-readable JSONL files. This allows for offline analysis of agent behavior while maintaining low-latency, in-memory streaming for the live dashboard.

To configure Claude Code hooks for session tracking, you launch the dashboard and point your hook configurations to the specified PAI_DIR directory. This setup captures deterministic agent events and streams them to the frontend at localhost:3000.

Yes, you can filter agent events by specific agent, session, and event type for focused debugging. The observability dashboard provides event filtering capabilities to isolate individual session flows and tool interactions within complex multi-agent workflows.