Start with the run
AI agent observability starts with a durable run record, not a dashboard chart. The record should answer which immutable agent version ran, how it was triggered, its current or terminal status, when it started and ended, and how many attempts occurred. Without the version link, later prompt or tool changes make historical results ambiguous.
Run status also needs consistent semantics. Queued, running, succeeded, and failed are operational states, while an error classification explains whether retrying was appropriate. A transient timeout and an invalid model request may both produce failures, but they demand different action. See error and retry monitoring.
Trace the work
A single duration cannot explain an agent that calls tools. Hierarchical spans show the root run, model-generation rounds, and bounded tool calls as parent-child work. This makes it possible to see whether latency came from the model, GitHub, queue retries, or another execution step.
Trace data should remain deliberately safe. Operational identifiers, statuses, durations, token counts, byte counts, and truncation flags are useful. Complete prompts, tool arguments and results, raw provider responses, credentials, and final model output do not belong in general-purpose logs or traces. Runovio keeps those boundaries separate; consult the security overview.
Read usage in context
Tokens, estimated cost, and latency are related but not interchangeable. Input, cached-input, and output token categories can have different rates. An estimate is only available when the selected model and required usage categories exist in the reviewed pricing snapshot. Unknown rates should remain unavailable rather than silently becoming zero.
Latency should be compared with attempts and trace shape. A run with one model call differs from a run that needed several serial tool rounds. Cost should likewise be read against the exact version, model, and pricing snapshot. These signals support investigation and capacity decisions; they are not a provider invoice or a guarantee of model quality.
Runovio places these signals beside each run so an operator can move from summary to trace without losing context. Continue with the traces guide or the AI agent observability use case.