Tracing What Agents Change When Required Information Is Missing

Agents that use API tools can change back- end records, so final task success can hide whether a consequential write was justified. The risk appears when a lookup needed for that write fails, yet the agent later changes state without visible support. APIFault makes this case reproducible by injecting re- alistic service faults into task-critical reads while keeping the original task, tools, user simulator, and reward fixed. It audits the trace after the first failed required read, cred- iting support only from successful tool ob- servations or protocol-authorized user fields. In audited Tau2-Airline slices, attempts to recover from faults can preserve completion while still producing unsupported backend commits. Policies that stop, hand off, or re- quire fresh verification reduce those commits, but shift cost to autonomy, escalation, or fall- back. For agents with state-changing tools, evaluation should trace what changes after required information is missing and check whether each change was supported at action time.