The real unlock with claws isn't the LLM itself, it's the orchestration layer that lets you chain tools together with state management between steps. I've been building multi-step automation pipelines (not code-related) and the hardest part is never the AI inference - it's handling failures gracefully, caching intermediate results, and knowing when to ask a human vs retry. The OTP/approval gate discussion in this thread is exactly right. The permission model needs to be as thoughtfully designed as the agent logic itself.