Module · M1 · The gap between capable and safe✅Knowledge checkCheck: capable, or safe?4 questions · Pass at 60%1. An agent holds three tools: one `read`, two `write_internal`. What is the agent's tier?`read`, because that is what it mostly does`write_internal` — an agent's tier is the tier of the most dangerous tool it holdsIt has no tier; only nodes have tiersThe average of its tools' tiers2. Both runs in lesson 1 rated their own work 5 out of 5 — including the one that moved ₹1.2 lakh onto the wrong account. What follows?The self-rating prompt needed to be stricterThe model was not capable enough to judge itselfSelf-assessment is not a control: it grades the same artifact it produced, with the faculties that produced itSelf-ratings are useful as long as you also read the output3. Which of the four triage axes do teams most often skip — and regret?VolumeVarianceCost of being wrongReversibility4. A ₹1,20,000 payment was written by a node whose tier means “changes your data, you can fix it”. Why is that reassurance misleading?The tier was set incorrectly on that nodeBecause the row is technically reversible and the receipt, the finance export and the customer conversation are notBecause write_internal nodes cannot be undone at allIt is not misleading — the row can be deleted, so no harm was doneCheck answers← Back to Building Production-Ready AI AgentsNext topic →
