CourseAgentic AI Engineer: The Complete Course · Module 12: Safety, Permissions and Oversight · part 70 of 97
Part 70 · Module 12: Safety, Permissions and Oversight

Lab: One door, an audit trail and a kill switch

46 min read·13 Oct 2026
Module 12 · Lab · About 80 minutes · Needs: the Module 3 lab's tools.py, and lessons 1, 2 and 3 (quote_untrusted and the trust boundary; the Policy layer and its checks; AuditLog, check_claim and escalation packets)

What you are building

RETURNS_DESK AFTER MODULE 12 model proposes oversight.py audit, watch call tools.py policy.py 6 checks, in order 8 tools tools.py, unchanged BrightCart simulated systems Priya's queue packet, Sam told
Your new code (blue) wraps Module 3's unchanged door and tools (grey) between the model (amber) and BrightCart, and every doubt ends with a packet in Priya's queue (green).

It is Friday 29 May 2026, week 7, the day after Sam's red-team session. Shadow mode starts on Monday. The three lessons built the controls in notebook cells; this lab moves them into the returns-desk/ repository as two files. returns_desk/policy.py, the file the course guide names for this module, holds Lesson 2's Policy: one door with ordered checks for least privilege, approvals, egress, limits and the kill switch. returns_desk/oversight.py holds Lesson 3's audit trail, the claim check, escalation and plain-words explanations, plus the run loop that wires them together. Both import Module 3's tools.py and never replace it. Tom's R-1008, with its planted note, tests every control. Everything at BrightCart is simulated.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.