Documentation
Docs
Debugging

Debugging Distributed Agents

Pinpoint why an autonomous system failed without searching through 50 different log streams.

The Observability Gap

When an agent-driven workflow fails in production, the "State" is scattered across multiple pods. Attaching a debugger to a single container is useless because the failure likely originated three steps ago in a different service.

The Trace ID: Your Unified Lever

Every user interaction starts with a Trace ID. This ID is cryptographically propagated by the Consonant Relayer across every agent call, tool execution, and database query.

Multi-Pod Log Tailing

bash
cons logs --trace-id d8f2-44a1-8e9c

This command automatically aggregates logs from all agents that participated in session d8f2-44a1-8e9c, sorted chronologically.

Waterfall Debugging

The cons trace command reconstructs the execution timeline. It allows you to see exactly where "Latent Hallucinations" or "Component Timeouts" occurred.

bash
cons trace d8f2-44a1-8e9c
[0ms]     SYSTEM: Goal Received
[120ms]   PLANNER: Plan generated (Step 1: Auditor, Step 2: Coder)
[150ms]   AGENT:auditor -> REQUEST: Check security of main.py
[1200ms]  AGENT:auditor <- RESPONSE: "File is safe" (tokens: 502)
[1250ms]  AGENT:coder   -> REQUEST: Refactor main.py
[15000ms] AGENT:coder   <- ERROR: Context window exceeded
Hallucination Detection

Consonant's sidecar uses small, specialized "Verifier" models to check agent outputs in real-time. If an output violates a predefined schema or policy, it is marked as a "Trace Fault," alerting you before the invalid data reaches your core systems.