Object Commitment as a Diagnostic Pressure Point in Grounded Planning

We investigate the role of explicit object commitment in non-linguistic agent planning within apersistent filesystem-based environment. To isolate its effect, we compare two agent variants thatshare identical perception, world models, and effect predictors: an action-only variant that resolvesobject arguments via external heuristics, and an object-centric variant that must select objectsinternally during planning.The action-only variant achieves consistently high task success; however, this performance masksbrittleness by avoiding object-level decision-making. In contrast, requiring explicit object commitment leads to lower success rates but exposes structured and repeatable failure modes—such aspersistent selection of incorrect objects—that are entirely absent in the heuristic-resolved variant.These diagnostic failures arise because the underlying dynamics and effect models are deliberatelyobject-agnostic and cannot distinguish consequences across specific objects, causing small predictionerrors to compound during planning. Moreover, we show that increasing training data exacerbatesthese failures rather than resolving them, revealing a representational bottleneck rather than anoptimization deficit.Our results demonstrate that apparent success in action-only planning can conceal fundamentallimitations in global abstractions, and that explicit object commitment provides a necessary lens fordiagnosing grounding and affordance challenges in grounded agent systems.