Understanding Agent Behavior
Foundry uses OpenTelemetry standards for distributed tracing, integrated with Azure Monitor Application Insights.
What Tracing Captures
- LLM calls — Model, tokens, latency, response
- Tool invocations — Which tools were called, with what arguments
- Agent reasoning — Decision chains and state transitions
- Errors — Failed calls, timeouts, content filter triggers
Setup in Code
from azure.monitor.opentelemetry import configure_azure_monitor
# One line to enable full tracing:
configure_azure_monitor(
connection_string="InstrumentationKey=xxx;..."
)
# All subsequent SDK calls are automatically traced!
Viewing Traces
Traces are viewable in two places:
- Foundry Portal → Tracing — Quick inspection of agent runs
- Application Insights → Logs — Advanced KQL queries for deep analysis
🎯 Pro Tip: Always enable tracing in production. When an agent fails, traces show you the exact reasoning chain that led to the failure — invaluable for debugging complex multi-step workflows.