AI Product Manager Zhongzhiliao: A Coding Agent Acceptance Design Where No Evidence Means No Pass
AI Product Manager Zhongzhiliao shares the design of dsh-acceptance, a tool built on DeepSeek Harness to address the lack of independent verification when Coding Agents report completion. He proposes anchoring acceptance criteria to original requirements, decomposing them into checkable pass conditions, recording locatable evidence, and using five states—pending, pass, fail, unknown, blocked—to preserve uncertainty, explicitly refusing to convert absence of evidence into either pass or fail.
ImportanceMaterialEvidenceE3 inspectableWrite-upQuick
As Vibe Coding becomes widespread, Coding Agents can complete multiple features at once and report "done." AI Product Manager Zhongzhiliao faced the question: how do you confirm the Agent actually did it right? If a second Agent only reads the development summary without executing real verification, the result may just be another plausible conclusion.
He developed dsh-acceptance on DeepSeek Harness (DSH). The core idea is not to make AI do more things, but to transform an AI Coding delivery into an item-by-item checkable, evidence-supported, historically recordable acceptance process. The design starting point shifts the question from "should we arrange a second Agent" to "how to ensure every acceptance conclusion has clear standards and verifiable evidence."
The design follows five steps. First, the system preserves original task requirements and links each acceptance item to the corresponding task contract, distinguishing original requirements from later supplementary suggestions, preventing acceptance criteria from being replaced by the auditee's self-description. Second, vague requirements are converted into explicit acceptance conditions—for example, breaking down "health check interface works normally" into HTTP status code 200, response status field value of ok, and verification method of issuing a request and checking the actual response.