LLMs generate confident-sounding probabilities because RLHF rewarded the pattern, not because they did statistics. I stopped trusting them and added a structural gate instead — no claim is acted on without an explicit, observable condition that would prove it false.
The dream is one agent that spawns a team. The failure mode is one agent that spawns agents that spawn agents. Spawning has to be a privilege — capped by depth, breadth, and role.
MoFlo gives Claude Code a memory, learned model routing, and repeatable workflows — local-first, no API keys, one npm install.
Approve every agent action and you'll rubber-stamp the one that matters. Approve none and you're the bottleneck. The real line isn't risky-vs-safe — it's whether it changes state.