Test the generality of system-level behavioral control

Investigate whether the finding that agent behaviors respond more strongly to system-level design than to user prompting generalizes beyond verification behavior to other behavioral dimensions.

Background

The paper evaluates configurable agents across information, reasoning, verification, budget, and model axes, and uses a trajectory-based taxonomy to analyze agent behavior. Its clearest system-level intervention is an oracle verification tool, which substantially increases reference-based verification and improves outcomes, whereas prompting the agent to verify has little effect. The authors explicitly caution that this conclusion was demonstrated for only one behavioral dimension and requires validation across additional dimensions before being treated as a general principle.