Skip to main content
Agentic operations is a people change wearing a technology costume. This chapter covers the shift from doing operations work to supervising it, how each operations role changes, and the four-step trust ladder for handing work to agents. Section 7.5 maps each step of the ladder to CloudThinker settings.

​
7.1 From executor to supervisor

“Human in the loop” means a person sits inside the execution path: every step waits for them. “Human on the loop” means a person supervises: they set intent and policy, approve consequential actions, audit outcomes, and step in by exception. Think of a pilot hand-flying a plane versus a pilot managing the autopilot. The second pilot isn’t less important. They’re responsible for more of the aircraft with less manual work, and they take the controls when judgment matters most. Teams that do well redesign roles on purpose. Teams that don’t end up with engineers who rubber-stamp approvals they no longer understand.

​
7.2 How roles change

Two new jobs appear. Treat both as career paths, not side duties.
  • Agent operations engineer. Owns the health of the agents themselves: instructions, connections, skills, memory, evaluations, and cost.
  • Autonomy policy owner. Often a senior SRE or engineering manager. Decides which kinds of action move up the trust ladder, and settles disputes when an agent and a person disagree.

​
7.3 Trust is built in increments

Engineers come to trust an agent in a predictable order. Skipping a step backfires. The 70% and 80% bars match the pilot exit criteria in §9.2. In CloudThinker, on-call rates each root-cause analysis under Was Resolve’s diagnosis right? on the incident, and only the workspace Admin role can record that rating. Give your graders the Admin role, or have an Admin record on-call’s verdicts. The failure mode is consistent: skip steps and you ship an agent that is confidently wrong at 3 a.m. One such incident can set an agentic program back by months. Take the steps in order and autonomy compounds. A practical test for step 1: give the agent an incident your team has already closed, and compare answers.
Replay several closed incidents this way. If it gets the root cause right on at least 70% of them, with evidence that matches your postmortems, move to step 2.

​
7.4 Measure the time you get back

The payoff of the supervisor model is senior time moved from toil to prevention work. Measure it, or it won’t show up in the business case:
  • On-call load: pages per engineer per week, with after-hours pages counted separately, from your paging tool (§8.2).
  • Toil hours: hours of triage and evidence gathering per week, from your ticket system or a two-week time study (§8.3).
Capture both before the pilot and again at each step of the ladder.

​
7.5 The trust ladder in CloudThinker

In CloudThinker, reads always run and every write passes through Approvals. Each step of the ladder is a setting you can see and change. New workspaces start on Auto · Rules + AI · Ask me. There, accepted Rules run, and AI judgment can also let a write no Rule covers run by itself when it judges the write safe and within what you asked. That puts the default past step 3 for those writes, so switch to Manual for step 1. Move to Auto · Rules + AI · Ask me only once your team trusts the agent’s judgment on your systems. Next: Chapter 8 · Measuring What Matters shows how to prove the agents are working, with the KPIs and evidence that justify each step up the ladder.