Back to AI/ML Overview
🔭 Knowledge Site · Observability for Agentic Software · Logs, Traces, Audit

Observability EngineeringA log records. A metric summarizes. A trace connects.

An agent can tell you what it did, and be wrong about it. This site starts from one line printed to a terminal and climbs, one idea at a time, to structured logs, metrics, traces, audit events, redaction, service level objectives, replay and cost, asking one question at each step: what does this record let me prove about what happened?

🎯The problem

A roof-assessment call times out, the agent retries, the run is checkpointed, and an underwriter decides hours later. Weeks after, someone asks why the quote was declined. The answer is spread across gateway logs, model spans, a provider timeout, a retry, a checkpoint and an audit event, and only the right identifiers join them into one story you can trust.

🔑The question every page asks
What does this record let me prove about what happened? Logs, metrics, traces and audit events each answer a different question, and each is followed through the same synthetic insurance run.

🗺️What is inside

Logs, metrics, traces and events

From stdout and structured logs to spans and span trees; what each signal answers, and the identifiers that join them across HTTP, queues, tools and models.

GenAI telemetry and OpenTelemetry

Token usage, model calls, retrieval and tool steps; the GenAI semantic conventions as read today, with their development status stated.

Audit, privacy and backends

Audit logs against diagnostic logs, redaction, access, retention and sampling, and the trade-offs of the common log and trace backends.

SLOs, replay, incidents and cost

Objectives and alerts worth waking for, replay and its limits, incident response, golden-case promotion, and cost per run.

🖱️How to read it

  • New to observability: read in order. Each chapter starts in plain English with one tiny example.
  • Experienced: jump to tracing, GenAI telemetry, audit logs or the insurance lab. The rigorous part sits in folded blocks.
  • Fourteen interactive labs, among them a log explorer and a trace waterfall, each with a plain table next to it.
  • The example is synthetic and its agent is a scripted stand-in. Costs are invented units. Results show how the mechanism works, not how good any model is.