Tracing without the platform

You probably don't need a tracing platform on week one. You need disciplined logs. Tracing platforms (Langfuse, Helicone, Arize, Datadog) become high-leverage when traffic is real and the team is multiple people. Before that, JSON logs in S3 will tell you everything you need to know.

What to log per run

• run_id, user_id (if applicable), timestamp
• Input — the trigger that started the run
• Every model call with its prompt, response, tokens, latency, cost estimate
• Every tool call with arguments, result (or error)
• Final output and exit reason (done, budget exceeded, error)

What to skip

• Secrets — redact before logging. Always.
• PII you can avoid — log the smallest amount that helps you debug.
• Anything 'in case it helps later' — bloat kills usefulness.