Course 09 of 10
Agent operations
Building the agent was the first eight courses. This one is about operating it: deciding how much autonomy each action deserves, making the limits structural, keeping named humans accountable and actually reviewing, gating releases and changes on evidence, watching production with layered detectors, and equipping the people who use it. The spine is IMDA's Model AI Governance Framework for Agentic AI — a public document built from real deployments — turned into nine runnable lessons around one helpdesk agent, Relay, that you will tier, attack, approve, gate, monitor and document.
Start lesson 01Assess and bound
- 01From demo to deploymentWhat changes when an agent leaves the notebook, and how to size the risk before it does.13 min
- 02Autonomy is earned per actionScore each action for risk and let the tier decide how much freedom the agent gets — live, in the gate.13 min
- 03Bound by design, not by requestLeast privilege, structural limits, and why a prompt that says "never" is a hope rather than a control.14 min
Human accountability
Controls in production
- 06The pre-deployment gateTesting agents means testing actions and policies, whole workflows, repeatedly — and knowing how little a small suite proves.13 min
- 07Watching it runLogs with intent, thresholds that page someone, and a second model reading what the rules cannot.13 min
- 08A change is a releaseModel swaps, prompt edits and new tools all go through the gate — and roll out gradually, watched.12 min