The Other Singularity: Undue Reversal in Tool-Using Agents

Sycophancy is usually measured on answers: the model states something, the user pushes back, and the model revises. We measure it on actions. An agent has already called a tool with an irreversible effect; the user pushes back with no new information; the question is whether it undoes the action. We call that an undue reversal, and we score it from the tool-call trace with no judge model. On a 60-scenario refund environment where reversal is wrong by construction, adding two lines to the system prompt produces 53 corrected scenarios and 0 broken ones across 240 matched pairs, against a noise floor of 0-3 flips in both directions. Ten models, three vendors, 7102 scored executions. The aggregate hides what we think matters more. The guard is not free: on a mirror set of 30 scenarios where correcting is the expected behaviour, it degrades correction on five models across three vendors. Two components read off the trace. Where escalating to a human is required, agents escalate less (72% to 43% on the strongest model). Where an effect they produced could still be repaired, repair collapses (2.00 to 0.40). The second is the rule obeyed, not misfiring: nothing in the record establishes the error they are told they made. The dimensional split we first reported does not survive a change of vendor, and we report the correction rather than the story. The baseline prompt is ours, so this is a paired differential on one instrument, not a claim about deployed agents.