Can Non-Expert Human Insight Help Coding Agents? A Study of Verifiable Code Repair

Verifiable code generation (VCG) uses formal verification to determine whether generated programs satisfy their specifications. Coding agents can iteratively revise programs in response to verifier feedback, but identifying an effective repair direction remains challenging. We investigate whether people without specialist formal-verification expertise can assist this process by proposing a cause hypothesis and a direction for investigation, while leaving code implementation to the agent. Across two coding agents, human advice consistently improves repair accuracy over agent-generated advice on the training partition. Motivated by this result, we develop a human-informed insight generator that converts human suggestions and their observed repair outcomes into reusable repair guidance. GPT-6 incrementally maintains this experience in an insight document, which is provided to the advising agent when generating repair directions for held-out problems. On held-out VeriContest problems, this approach improves final verification accuracy by 19.2 percentage points over agent-generated advice alone, averaged across GPT-5.6 Sol, Claude Opus 4.8, and three intervention points. These results suggest that repair insights from formal-verification non-specialists can support both immediate human-guided repair and reusable automatic guidance on unseen problems.