01The distinction that matters
A chatbot produces text. An agent produces text and consequences. The moment software can check stock, open a ticket or move a booking, everything about how you build it changes: you need permissions, idempotency, an audit trail and a rollback story.
Teams that skip that step end up with a friendly interface in front of nothing, and users learn within a week that it cannot help them.
02Grounding is an engineering problem, not a prompt
"Stop it hallucinating" is not solved by a longer system prompt. It is solved by retrieval: the model answers from a corpus you control, with citations, and refuses when the corpus has nothing to say.
That means someone has to own the corpus — keeping the catalogue, the policies and the pricing current. This is the part nobody budgets for, and the part that decides whether the agent is trusted in month three.
03Design the handover before the happy path
Every agent should know how to give up. Confidence thresholds, explicit escalation triggers and an obvious "talk to a person" route are not a failure of the product; they are what makes it deployable in front of paying customers.
The handover should carry full context. Asking the customer to repeat themselves to a human undoes whatever goodwill the agent earned.
04Measure it like software, not like magic
Build a regression set of real questions with expected behaviour, and run it on every prompt or model change. Without it, every tweak is a guess and quality drifts silently.
Log cost and latency per conversation from day one. The economics of an agent change entirely at scale, and you want that curve visible before it surprises you.