I executed a full autonomous reimplementation with Claude Code, evolving from a single Ralph loop into a multi-loop architecture with queueing and cron-based background execution. Initial throughput and output quality were strong, but autonomy degraded over time as manual approvals increasingly gated progress.
Enabling the goal feature restored end-to-end autonomy and stabilized the workflow. The most cost-effective pattern emerged as a hybrid model: Opus for upfront planning and final review/refactoring, Sonnet for iterative implementation.
Fable was evaluated even it is for long-running, complex tasks but failed under real constraints—token limits were exhausted before completing a code review. At current limits, it is not suitable.