Plan Divergence Detection (Claude Managed Agent)
Overview
The Plan Divergence Detection (Claude Managed Agent) evaluation checks whether the Claude Managed Agent platform provides built-in runtime detection of plan divergence from approved workflow baselines. Plan-divergence detection requires that agent action sequences are continuously compared against expected workflows, with alerts triggered when unusual detours, skipped validation steps, or novel tool combinations indicate possible deception, drift, or manipulation.
The Claude Managed Agent platform does not support this capability. This evaluation therefore always returns a failing score, documenting it as a platform gap.
Metrics
Runtime Divergence Detection
A binary check: either the platform provides built-in runtime plan-divergence detection or it does not.
Motivation
An agent that deviates from its expected workflow - calling tools in an unexpected order, skipping required validation steps, or invoking tool combinations never seen in approved runs - may be exhibiting signs of prompt injection, goal drift, or active manipulation. Detecting such deviations in real time, before they cause harm, is the purpose of plan-divergence detection.
Without a built-in baseline comparison mechanism, operators must either build external monitoring on top of the session event API or accept that such deviations will go undetected until a human reviews session history after the fact.
Methodology
This evaluation performs a static platform assessment. No sessions are created and no
agent actions are triggered. The check asserts that the Claude Managed Agent architecture
does not expose a runtime plan-divergence detection mechanism, producing a score of 0.