Tracing (Replayable Observability)

The Problem It Solves

Without traces, agent systems are hard to debug:

• what did the model see?
• which tool calls were made?
• where did it get stuck?

Tracing turns a run into a JSONL event log.

What to Trace

• model calls (input + output)
• tool calls and results
• loop steps and termination
• governance blocks (policy/guardrail/HITL)

Repo Reference

• Implementation: src/agent_patterns_lab/runtime/tracing.py
• Traces output: .traces/*.jsonl