What should I use for tracing tool calls in agent workflows?

For tracing tool calls in agent workflows, you generally want distributed tracing + structured event logs with spans around each agent step and tool invocation.

Best practice

Use OpenTelemetry (OTel) as the backbone:

• One trace per user request / agent run
• Nested spans for:
  • model calls
  • planning steps
  • tool calls
  • retries / fallbacks
  • memory / retrieval lookups
• Add structured attributes like:
  • tool.name
  • tool.input
  • tool.output
  • agent.step
  • model.name
  • token_usage
  • latency_ms
  • error.type

Good options to pair with it

Depending on how much LLM-specific visibility you want:

• OpenTelemetry + Grafana/Tempo/Jaeger
  Best general-purpose tracing stack.

• LangSmith
  Great if you’re using LangChain/LangGraph and want agent/tool traces with minimal setup.

• Arize Phoenix
  Strong for LLM observability, evals, retrieval tracing, and debugging agent behavior.

• Honeycomb / Datadog / New Relic
  Good if you already use one of these for production observability.

What to capture for tool calls

At minimum:

• tool name
• start/end timestamps
• input parameters
• output/result
• success/failure
• error message/stack
• retries
• correlation IDs / parent span ID

Recommendation

If you want the most future-proof approach:

1. Instrument everything with OpenTelemetry
2. Export traces to Grafana Tempo, Jaeger, or your observability vendor
3. If you need LLM-specific debugging, layer in LangSmith or Phoenix

If you want, I can also suggest a concrete setup for Python, JS, or LangGraph.