Implement multi-step workflows with enforcement and handoff patterns

Strengthen the system prompt so the required step is unmistakably mandatory.

Prompt compliance is probabilistic. At any real volume a small failure rate becomes routine, and in a payments workflow those failures are misidentified accounts and wrong payments.

Where a step must always happen, make the next call structurally impossible until it has.

Add few-shot examples that always show the correct order.

Examples raise the probability of compliance without ever guaranteeing it — the same class of control as the prompt, wearing different clothes.

Few-shot shapes judgement and format. It does not produce deterministic ordering.

Add a classifier that enables only the tools appropriate to each request type.

Right primitive, wrong layer: it controls which tools are available, not the order they are used in. Within an enabled set the ordering is unconstrained again.

Availability is not sequence. Gate the prerequisite, not the toolbox.

Escalate by handing the human the conversation transcript.

The receiving agent usually has neither access nor time, so the customer re-explains everything and the escalation loses whatever the agent had already established.

Hand over a structured summary — identifier, root cause, amount, recommended action — that stands on its own.

Answer a message raising three problems as a single request.

One of the three gets a real answer and the others are acknowledged and dropped, which reads as a reply and is a partial resolution.

Split into distinct items, investigate them in parallel against shared context, then synthesise one unified response.