Observability
Transparency in the Age of Agents
Debugging AI agents can be a nightmare of black boxes and stochastic behavior. HotPlex solves this by building Observability directly into the core engine. Every thought, action, and failure is recorded and visible.
The Three Pillars of Agent Visibility
1. 🔍 Real-time Trace Logs
Every session in HotPlex generates a structured trace. You can follow an agent's logic in real-time via the CLI or the Admin Dashboard:
- Trace Level: See exact token counts and model latency.
- Action Level: Track every tool call and its raw result.
2. 📊 Performance Metrics
HotPlex exposes standard Prometheus metrics, allowing you to monitor the health of your agent infrastructure:
- P99 Latency: Measure how fast your agents are responding.
- Error Rates: Identify which hooks or tools are failing most frequently.
- Token Usage: Track model costs in real-time across your entire fleet.
3. 🛡️ Audit Trails
For enterprise compliance, every interaction is recorded in an immutable audit log:
- Who: Which user or system triggered the session.
- What: The exact input, thought process, and output.
- Where: The ChatApp or API endpoint used.
Integration with External Tools
HotPlex is designed to play well with the modern observability stack:
- OpenTelemetry: Native support for exporting traces to Jaeger, Honeycomb, or AWS X-Ray.
- Grafana: Pre-built dashboards for monitoring agent health and performance.
- ELK Stack: Stream structured logs directly to Elasticsearch for deep analysis.
bash
# Export traces to an OTLP collector
hotplexd run --otel-endpoint "otel.internal.yourcorp.com:4317"