Alibaba Cloud's Indonesia community published on October 27 an end-to-end case study of a support-ticket and refund assistant built with the Qoder Agent SDK. The worker starts a session per case, loads a local skill describing the review procedure, and calls two scripts against the existing application: one to read case details with evidence, one to append a review note without changing case status or version. The guardrails sit in the backend rather than in the prompt. An agent decision is committed only when its case version matches, confidence is at least 0.9, the amount is no more than 250 dollars, and policy findings are complete. Approvals additionally require a full refund, matching identity, enough photos and no pending seller-response window; automatic rejection is limited to a verified expired return window; everything else goes to a person. The author notes that a photo caption is not proof a label is visible, and that such distinctions belong in the skill text where they can be revised.
Our takeThe architecture is the argument: the model proposes, the service disposes, and the checks that matter are version equality and completeness of findings rather than the confidence number itself, which is why a high score alone does not unlock an approval. Keeping the procedure in a skill file makes the policy reviewable without redeploying the worker. The limit is coverage: rules this crisp exist for refunds, and the same pattern will look far weaker on cases where the policy finding is itself a judgement.