CourseAgentic AI Engineer: The Complete Course · Module 12: Safety, Permissions and Oversight · part 68 of 97
Part 68 · Module 12: Safety, Permissions and Oversight

Lesson 2: Controls the model cannot talk its way past

33 min read·13 Oct 2026
Module 12 · Lesson 2 · About 60 minutes · Needs: Modules 3 and 5 (tools, loop, approvals), Module 10 Lesson 3 (dry runs)

The situation

It is Friday 29 May 2026, week 7. Yesterday Sam red-teamed the agent with R-1008, Tom's message with the planted note. Shadow mode starts on Monday.

Sam wants one place to read every rule: "Your guards are scattered across the tools, the systems and the loop." Marta wants a ceiling on what a bad day can cost. Priya wants a button that stops every run. And Dev has just added fetch_page, so the agent can read help-centre pages.

This lesson builds that place: a Policy layer between the model's tool calls and the tools. R-1008 tests every control; Aisha's R-1001 checks that honest work still gets through.

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.