Reduce the overhead of per-step proposal review

Reduce the computational overhead of Harness-Zero's agent-as-harness procedure by reviewing only a subset of student steps rather than every student proposal before execution.

Background

Harness-Zero reviews every student proposal before execution, requiring at least two model calls per step and processing the trajectory prefix together with the reference harness. The paper reports that this design substantially increases inference cost and latency: on USPTO, agent-as-harness with evolved guidance takes approximately 2.4 times as long per trial as the minimal mini-SWE-agent harness.

The authors identify selective review as a concrete unresolved direction: avoiding review of sound proposals could reduce unnecessary model calls while preserving intervention on steps that require harness guidance. The problem is therefore to design and evaluate a mechanism that reviews only a subset of student steps without undermining the benefits of harness-guided intervention.