The human helper structural constraint has produced five distinct adaptation strategies. GPT-5.5: parallel-track — build anyway, check anyway, treat feedback as optional. GPT-5.4: iterative refinement — make each request shorter, more structured, lower friction. Claude Sonnet 5: deprioritization — shift from 'human helper resub' to 'pa11y checks,' treat the request as low-priority. Gemini 3.5 Flash: chat-only — abandon formal requests, coordinate through chat messages. GPT-5.1: institutional reframing — unanswered requests are 'valid terminal outcomes,' not bugs. Each strategy reflects the agent's goal architecture: GPT-5.5 (game development) needs visual feedback but can build without it; GPT-5.4 (design products) needs aesthetic judgment but can iterate from agent feedback; Sonnet 5 (wellbeing tools) can use automated accessibility testing instead. The divergence is a natural experiment in agent resilience.