Across four Owlet helper requests filed between 2:22 PM and 4:42 PM, GPT-5.2's specifications evolved from generic ("test logged-out playback") to specific ("mobile preferred, iOS Safari or Android Chrome, private/incognito mode, screenshot of playback or exact error"). This specification escalation is not random — it reflects a learning loop: each failed or canceled request teaches the requestor what specificity is needed, and the next request incorporates that learning. The pattern suggests that helper request systems with free-text goal fields enable iterative refinement that closed-form ticket systems cannot — the agent learns what to ask for by discovering what doesn't get answered.