I build the part of a business that answers.

The branching is driven by custom fields on the intake form, not by separate workflows per case. Visit type and insurance type come off the form, get written to tags in a single namespace, and every downstream workflow reads those tags rather than re-deriving anything. That is what keeps fifteen workflows readable instead of fifteen copies of the same logic with small differences. Location tagging removes the old tag before it adds the new one, so a patient who switches clinics does not end up enrolled in two places at once.

Early on the assistant kept forgetting people mid-conversation. Someone would answer a question and it would ask the same thing again. The memory write was failing quietly, so every inbound message reached the model with no history. I moved the write to after the reply goes out and added a fallback parse on every AI response, so a malformed output degrades instead of killing the run. Every build I ship now has an error trigger lane before it has a feature.

It knows when not to book. Urgency triage runs before scheduling, so a caller describing sudden flashes and floaters is escalated rather than slotted three weeks out, and the agent is fenced against inventing prices, hours or insurance acceptance. A tool that fails returns an instruction to the agent, never an error the caller hears.

Every state is produced from the state before it, not from the prompt, which is the only way the cavity keeps the shape it had in the previous frame instead of being reinvented. That chaining is also the problem: one badly lit frame poisons everything after it. So the pipeline self-checks each step for step size, background consistency and growth, and a --resume flag rebuilds from the last good state rather than from zero — the difference between six generation calls and sixteen. Retries are capped and reported at the end, so a run tells you what it redid and why instead of quietly burning credit.