8. Agentic Workflow Design

Domain 3 (27%) — Agentic Architecture & Orchestration

Escalation: when and how — Decision logic and handoff content

Escalate when:

• Customer explicitly asks for a human
• Issue requires authority the agent doesn't have (policy exceptions, amounts above limits)
• Agent can't make meaningful progress

Warning

Don't use mechanical rules like "escalate after 3 failed tool calls" or "escalate when sentiment score exceeds threshold."

Handoff content: structured summary (customer ID, root cause, transaction IDs, amounts, recommended action). Not a raw transcript. Not just the original complaint.

Handling frustrated customers — correct sequence — Order matters

1. Acknowledge the frustration first. Don't silently investigate their account.
2. Ask one targeted question to understand the issue.
3. Then decide: escalate or resolve directly.

If the issue can be resolved and the customer still wants a human: offer the choice. Don't force either path. Don't process the return unilaterally.

Enforcement spectrum — From unreliable to deterministic

Method	Reliability
Prompt instructions	Lowest reliability. Non-zero failure rate. Soft guidance only.
Few-shot examples	Better than written rules. Demonstrations stick longer.
Tool-level validation	Schema constraints, parameter checks. Reliable for structure.
Orchestration-layer hooks	Highest reliability. For hard compliance rules: intercept tool calls, check conditions, block execution. Model is taken out of the decision entirely.