Why Deterministic Workflows Beat LLM-Powered Routing: The Claude Sub-agents vs. duckflux Case

⚡ Key Takeaways

• Claude Sub-agents delegate routing to LLM inference, which is unreliable for repeatable pipelines—LLMs forget steps, miscount iterations, and silently skip transitions 𝕏
• duckflux separates deterministic orchestration (YAML config) from creative execution (LLM work within steps), matching how real workflows actually operate 𝕏
• Not all pipeline decisions need LLM routing—step ordering, retry logic, and error handling should be deterministic config; only the work inside steps should use LLM reasoning 𝕏