The smartest part of my outbound system has the least power.
I built it that way on purpose.
It's the single agent that uses an LLM to write the actual outreach. It can draft. It cannot send. There is no path from that agent to a send channel at all. Its only output is a draft that sits in a queue, waiting for a deterministic quality gate and human sign-off.
People find that backwards. The LLM is the most capable piece, so why give it the least reach?
Because capable and predictable are not the same thing.
The LLM is the part most likely to hallucinate a detail or write something I would never put my name on. So it gets the tightest constraints, not the loosest.
The predictable work runs on plain deterministic code. Routing. Scoring. Rate limits. The kill switch. No model involved. The judgment goes to the LLM. Everything else stays boring on purpose.
Most of the skill, it turns out, is knowing where NOT to put the AI. Wiring a model up to do everything is easy. Deciding what it should never be allowed to touch is the hard part.
It can write. It cannot send.