HarnessQuotient: Counterexample-Guided Behavioral Quotienting for Self-Evolving Agent Harnesses

Self evolving agent harnesses accumulate tools, skills, specialist agents, memory operations, and runtime policies. As the harness grows, many capabilities become redundant in appearance, but similarity is not enough to establish that two behaviors are interchangeable. Procedures can agree on common cases and still diverge in a rare context that matters for future execution. Compressing such procedures into one semantic action removes a distinction the agent still needs, a failure we call distinction forgetting. We introduce HARNESSQUOTIENT, a framework that treats every proposed merge as a behavioral claim to be tested. Candidate pairs are identified from inexpensive structural and semantic signals. A semantic challenger then proposes conditions under which the pair may disagree, and a planner spends a finite verification budget on valid contexts that are most informative about those possibilities. Both capabilities are executed from matched state. A discovered difference becomes a persistent separating witness, while negative evidence updates a risk-controlled certificate. The system merges only when the remaining risk is sufficiently small and otherwise leaves the pair unchanged. Semantic classes require evidence against all existing members, and certificates are reconsidered when the harness changes. This view turns harness compression into a problem of preserving the right distinctions while removing the unnecessary ones.