Tracing and Observability¶
Every prompt, tool call, and model response an agent makes is invisible unless you record it. This subtopic is how you make a run reconstructable after the fact — and how those recordings become production monitors.
flowchart LR
J["Junior: record one run"] --> M["Middle: build a readable run tree"]
M --> S["Senior: treat the schema as a budgeted interface"]
S --> P["Professional: make trace quality an SLO"]
Levels¶
| Level | Guide | You are done when |
|---|---|---|
| Junior | Record one run | You can reconstruct a single run — every prompt, tool call, and response — from what you captured. |
| Middle | Build a readable run tree | You can instrument a multi-step agent as parent/child spans a teammate can read without you. |
| Senior | Treat the schema as a budgeted interface | You can design sampling and prod monitors that catch drift without capturing everything at full fidelity forever. |
| Professional | Make trace quality an SLO | You can set an org-wide trace standard with a retention and compliance policy teams don't each reinvent. |
Practice rule¶
If you can't answer "what was the exact prompt sent to the model on this run" from your logs, you don't have observability — you have print statements. Record the rendered prompt, not the template.
Related¶
- Evaluation Fundamentals — the metrics computed from what gets traced here.
- Debugging Agent Failures — the discipline of reading a trace to find the first wrong step.
- Cost and Performance — token and latency numbers pulled directly from spans.