Start
Lilium
Journal
Observability

Why agent observability needs lineage, not just logs

Long-running agents branch, delegate, retry, and change their own context. A flat log stream loses the shape of the work.

July 24, 20261 min read

Traditional logs are excellent at recording individual events. Agent systems add another question: how did this event inherit its context, tools, model configuration, and objective? Answering that requires the full run, not an isolated line.

#The run is the unit of understanding

Lilium treats every production run as a trace with durable parent and child relationships. Model calls, tool calls, sub-agents, and application spans keep their place in the same story.

A flat log stream can tell you that a tool failed. It cannot tell you which decision selected that tool, what the agent believed at the time, or whether a sibling branch had already succeeded at the same task. Those questions are structural, and structure is exactly what a line-oriented log discards.

#Lineage makes retries legible

A retry can look harmless in aggregate metrics and still be the source of a large latency or cost regression. When retries retain lineage, an operator can see which decision produced them and compare the arguments used each time.

That structure also makes failure clusters more useful. Instead of grouping only by error string, teams can group by the path through the agent that reached the error. Two runs that fail with the same message for different reasons stop being averaged together.

#Replay starts with faithful capture

A useful replay needs the prompt, tool boundaries, model settings, and relevant outputs from the original run. Capture should be configurable, redacted before storage, and explicit about what is missing.

With that foundation, a team can replay one branch, compare a new prompt, or test a model change without pretending the original production context never existed. That is the difference between debugging an agent and guessing at one.

If you are instrumenting for the first time, the quickstart gets a first trace flowing in a few minutes, and the OpenTelemetry guide covers teams that already run a collector.