Design a scoped, tool-using AI agent for a specific task

Help me design an AI agent for a specific, scoped task — not a general-purpose assistant.

Task the agent should perform: [describe precisely, e.g. "triage incoming support emails, categorize them, and draft a reply for simple cases, escalating anything about billing or refunds to a human"]
Tools/APIs it will have access to: [list them, e.g. "read email inbox, send draft (not send directly), look up customer record by email, create a support ticket"]
What it must NEVER do without human approval: [list explicitly, e.g. "never send an email directly", "never issue a refund", "never delete data"]
Model/framework: [e.g. Claude API with tool use, a specific agent framework — name it]

Design:
1. A system prompt that clearly scopes the agent to this task only, states its available tools and what each is for, and explicitly states what it must escalate to a human rather than attempt
2. The exact tool definitions (name, description, parameters) it needs — no more, no fewer than required for the task
3. Explicit instructions for what the agent should do when it's uncertain (e.g. "if you're not confident about the category, ask for human review rather than guessing") — the goal is safe failure, not guessing to seem competent
4. A clear boundary between actions the agent can take autonomously vs. actions that require a human-in-the-loop approval step
5. Logging/observability: what should be logged about each decision the agent makes so a human can audit it after the fact
6. A short set of test scenarios, including adversarial or ambiguous inputs, to verify the agent escalates instead of guessing when it should

Do not design this as an open-ended chat assistant — every tool and instruction should map directly to the one task described above.