Large language model agents often act under specifications like instructions or schemas, but the same model both executes and judges completion. This creates two gaps: requirements may be understood but not executed, and agents may claim completion before requirements are met. On SkillsBench, across seven models, only 79.6% to 86.4% of extracted task directions were satisfied, while agent completion claims exceeded official pass rates by 28.7 to 37.9 percentage points.
The authors propose SpecHarness, a runtime that separates agent proposals from authoritative state. Compiling specifications into source-linked obligations, SpecHarness governs execution and finalization independently, validating verifiable requirements at runtime and leaving ambiguous ones advisory. Experiments on guideline-following and artifact-generation tasks support treating specifications as authority over compliant execution and completion.