Skip to content

Observe It

Logs, metrics, traces, analytics — visibility without PII storms — matched to the quality regime.

Make production (or staging) speak so you are not guessing. Instrumentation shape follows the system: service golden signals, web RUM/vitals, or LLM traces and quality scores.

Google SRE’s four golden signals (latency, traffic, errors, saturation) are the minimum for user-facing services. Interactive products also need field Web Vitals and UX error paths. Generative systems need hierarchical traces of model/tool/retrieval steps plus scores (online evals, user feedback) — traditional APM alone under-describes quality (Quality regimes). OpenTelemetry keeps instrumentation portable; Langfuse is this house’s default LLM observability/eval plane on top of that idea. Watch cardinality and PII in every regime.

When prod/staging is silent, quality is drifting, or readiness needs live evidence — after naming the regime.

  • Pick signals for regime A (incl. data freshness/correctness), B (vitals + errors), or C (traces + scores)
  • Invoke the skill; follow its safety rules
  • Keep one write owner; collect evidence before claiming done
  • Hand off to the next practice instead of boiling the ocean
  • Equate “no 5xx” with “good generations”
  • Flood logs with prompts/PII
  • Skip orientation when you’re lost
  • Spawn overlapping agents to “go faster”

04-evidence-over-vibes, 11-quality-regimes, Operate

See also the handbook Sources & grounding bibliography.

observe-it · companion langfuse when operating LLM traces/evals