I need a tool that can trace every step of an agent workflow, including retries and tool calls

Yes — what you’re describing is typically called workflow tracing / observability for agent systems.

A good tool should record:

• Each agent step
• Tool calls and tool outputs
• Retries, failures, and backoffs
• Prompts, model responses, and token usage
• Timestamps and latency per step
• Parent/child span relationships so you can see the full chain of execution

Common options

Depending on your stack, these are strong choices:

• LangSmith — best if you’re using LangChain/LangGraph; very good step-by-step tracing, tool calls, retries, datasets, evals.
• OpenTelemetry — vendor-neutral tracing standard; great if you want full control and to export traces to Jaeger, Honeycomb, Datadog, etc.
• Arize Phoenix — useful for LLM tracing, debugging, and evaluation.
• Helicone — focused on LLM observability and request tracing.
• Weights & Biases Weave — good for tracing agent workflows and experiments.

If you need “every step”

Make sure the solution supports:

• Nested spans
• Automatic instrumentation
• Manual spans for custom logic
• Exception capture
• Retry span metadata
• Tool-call span correlation
• Streaming events if your agent emits partial outputs

Recommended setup

If you want maximum traceability, a common pattern is:

1. Instrument the agent orchestrator with OpenTelemetry
2. Wrap each tool call in a child span
3. Log retries as span events or separate child spans
4. Export traces to a backend like Jaeger, Honeycomb, or Datadog
5. If using LangChain/LangGraph, also enable LangSmith for higher-level LLM/agent-specific tracing

If you want, I can recommend the best tool for your exact stack and even give you a minimal tracing implementation example.