Workflow enforcement and handoff

• Programmatic enforcement (hooks, prerequisite gates) is deterministic; prompt guidance is probabilistic.
• When compliance must be certain (identity check before financial operations), prompt instructions alone have a non-zero failure rate.
• A structured handoff includes customer details, root cause analysis, and recommended actions.
• Multi-concern requests should be split into distinct items and investigated in parallel with shared context.
• A gate blocks a downstream tool (for example process_refund) until a prerequisite (a verified ID from get_customer) has completed.

• Implement prerequisites that block downstream tool calls until earlier steps finish.
• Decompose multi-concern requests, investigate each in parallel, then synthesise one unified resolution.
• Compile handoff summaries (customer ID, root cause, refund amount, recommended action) for human agents without transcript access.

• "A strongly worded prompt is as reliable as a hook." It is not; a 2% miss on refunds is still unacceptable.
• "Escalation can just point the human at the chat." Humans often cannot see it and need the facts up front.
• "Handling one concern per turn is fine." It lowers first-contact resolution.

Financial or compliance steps enforced only through prompt wording, and vague handoffs with no customer ID or root cause.

• 1.5Agent SDK hooks
• 5.2Escalation and ambiguity resolution
• 1.1Agentic loops