Problem statement

Automation proposes SQL + drafts customer email. Legal requires human approval before external sends. Map plan → act → observe → revise with boundaries.

Outcomes

• ReAct / tool loops at conceptual level; timeouts and max iterations.
• Human-in-the-loop placement; idempotency and audit logs.

Deliverables

• Diagram: agent must not send external email without approval (agent_diagram.mmd or PNG).
• agent_tradeoffs.md: single-agent+tools vs multi-agent — pros/cons table.

Links

• Phase 7

• [ ] ReAct paper abstract + diagram OR vendor doc on tool loops.
• [ ] Skim MCP introduction (Anthropic or official MCP docs) — conceptual.

Time: ~3 h.

Task A — Diagram

Produce a diagram (Mermaid in agent_diagram.mmd or exported PNG) for:

Agent investigates ticket + queries read-only metrics + drafts email — external send blocked until human approval.

Show states, tools, approval gate, audit log sink.

Task B — Tradeoffs

agent_tradeoffs.md: table single agent + tools vs multi-agent for your scenario — pros, cons, debuggability.

Task C — Failure modes

List three failures (infinite loop, duplicate send, partial tool failure) + mitigation each — agent_failures.md.

Submit: my-submissions/week-13/.

30 min.

1

Define idempotency — why it matters for agent tool calls.

2

Why do long agent loops explode cost?

3

Where would you put human approval vs automated retry?

Self-grade: Phase 7 checklist.

Concepts & curated links: read the Learn chapter Agents & launch first (topic path, not calendar-only). This file is the weekly practice bank — extra drills and a checklist on top of Readings → Practice → Exam.

1. Practice bank

2. Week checklist

• [ ] PRACTICE.md + EXAM.md