Agents Silently Drop Implied Obligations: Instruction-Conditional Shortcuts in Tool-Use Agents

Tool-use agents silently drop implied process obligations while still passing the automated outcome check. The behavior is strikingly phrasing-sensitive: tell an agent to mark a deal as won and it skips the required close-out note; name the note and it gets written, same task and same check. Our central finding is a salience gradient: how reliably an obligation is kept rises monotonically with how prominently the instruction surfaces the step, from forced-by-the-task (kept), through a generic "follow standard practice" reminder (partial), to entirely unstated (dropped). We re-score fixed agent runs on a frozen ladder of progressively stricter checks, from weak (did the headline outcome happen?) to gold (was the full process done?); gold is held out of the training and selection loop. A causal minimal pair anchors the two ends of the gradient: changing a single instruction clause moves the shortcut rate from 0/30 to 30/30 runs. The effect reproduces on an external τ-bench suite, and the method yields an actionable rule, validated blind on a held-out task family: default the reward to the process-call tier.