Module 10 Activity

Scenario

The capstone: a reviewed multi-agent automation with the evidence that its controls work.

What you build

A complete crew with guardrails built into the tools, full instrumentation, deliberate failure-path tests, and a handover note that states its limits.

Steps

1. Choose a workflow and run it through the four checks. Write down why one agent would not do.
2. Build the guardrails inside the tools from the start - including requiring the approver as a parameter.
3. Instrument all five sections before the first real run: per task, per tool call, per run, human review, budget.
4. Write one test per control that constructs the failing condition, including the duplicate approval that manual testing cannot produce.
5. Measure against the manual process, including review time and both error rates.
6. Write the handover note with known limits, an owner, and next steps. Give it to someone and record every question they ask.

Evidence to hand in

• The four checks and the why-not-one-agent argument.
• The tools with their guardrails, and the approver parameter.
• The full instrumentation schema.
• One test per control, including the duplicate approval.
• The measurement including error rates.
• The handover note and the reader's questions.

Review checklist

• Guardrails are inside the tools, not in prompts.
• approved_by is a required tool parameter.
• Every control has a test that makes it fire.
• The duplicate-approval case is tested.
• known_limits states any metric that is worse than manual, with the trade.
• The note names an owner, and the reader's questions were recorded.

Mark complete