An AI agent ticked a checkbox correctly and the form rejected it anyway. Four safari-mcp releases later — react-select, Vue, Safari geometry, macOS code-signing — here is every layer of the stack that can tell a synthetic click from a real one.
Liquid syntax error: 'raw' tag was never closed
Eight stages, two human gates, four memory layers. The full state machine for agentic coding, with the alternatives I rejected.
1h 26m from prompt to merged PR for a real fintech webhook handler. $1.42 in LLM spend across architect + pm + 2 senior-dev + 5 reviewers. The full trail with caveats.