Skip to main content
Definitions matter, because “agent” is the most abused word in enterprise software. This chapter gives you four tools: a definition you can quote, a five-level autonomy scale for deciding what an agent may do on its own, a list of what agentic operations is not, and a five-question test for vendors.

​
3.1 A working definition

DefinitionAgentic infrastructure operations (AgenticOps) is an operating model in which AI agents run the core loop of operational work (detect a condition, analyze the cause, resolve the issue, validate the outcome) across cloud and on-premise infrastructure. They work under explicit policy that humans define. Humans supervise on the loop instead of executing in it.
An agent in this sense is not a chatbot with a runbook, and not a script with an LLM bolted on. A real operations agent has five properties. Use them as a checklist when you evaluate any tool: If a tool is missing any one of these, treat it as an assistant, not an agent.

​
3.2 The autonomy spectrum

Autonomy is not on or off. Mature agentic platforms expose it as a policy dial, usually set per action class (for example “restart a pod”) and per environment (for example production versus staging): Most organizations run several levels at once. Every action class starts low and earns its way up. A typical ceiling once the evidence is in: Move an action class up one level only when you have evidence it works at the current one. Use the Chapter 8 KPIs as the gate: an acceptance rate above about 95% for that action class and a flat or falling rollback and intervention rate. Never faster. Chapter 7 turns this into a step-by-step trust ladder.

​
In CloudThinker

CloudThinker doesn’t use L0 to L4 labels. Reads always run, and every write passes through Approvals in Approvals in the sidebar, Approval mode tab. Each level maps to a setting: New workspaces start on Auto · Rules + AI · Ask me, which lets AI judgment allow a safe write and asks a person about the rest. CloudThinker turns on Rule enforcement one workspace at a time. While the Rules step shows In preview, an accepted Rule only records what it would have decided, so the Approval mode is the control that holds today.
How the hyperscalers ship the dial
  • Gemini Cloud Assist investigations run root-cause analysis and report observations. They don’t change your resources (L1).
  • AWS DevOps Agent investigates incidents and hands you a mitigation plan to carry out (L1 to L2).
  • Azure SRE Agent has a Review mode where actions wait for your approval, and a more autonomous mode for actions you allow (L2 to L3).
All three start an agent with little trust and let it earn more.

​
3.3 What agentic operations is not

“Agent washing” is common enough that Gartner has named it: vendors rebranding assistants, chatbots and RPA as “agents” without real agentic capability. In a June 2025 press release, Gartner estimated that of the thousands of vendors claiming agentic AI, only about 130 were real. Knowing what agentic operations is not is your best defense as a buyer:
  1. Not a chatbot over your dashboards. Chatting with your telemetry is a feature, not the operating model. If a human has to read the answer and then go do the work, you are still in Gen 3 (see Chapter 2), whatever the marketing says.
  2. Not lights-out operations. No credible practitioner advocates removing humans. The target is human leverage: one engineer supervising work that used to take a team.
  3. Not a replacement for engineering discipline. Agents amplify the environment they are given. Weak observability, absent IaC, and undocumented systems produce weak agents. Garbage context in, garbage autonomy out.
  4. Not one giant model that does everything. As the next chapter shows, production systems are converging on orchestrated teams of specialists, not monolithic super-models.
The five-question vendor testAsk any “agentic” vendor:
  1. Can the system execute a remediation end-to-end, or only recommend?
  2. Does it verify its own outcomes and roll back on failure?
  3. Can autonomy be set per action class and per environment?
  4. Does every action carry a full, immutable reasoning trail?
  5. What were its rollback and intervention rates in its last three production deployments?
A real platform answers all five with evidence, not a roadmap slide. Agent washing usually fails by question two.

​
3.4 The scope of operational work agents can own today

Use this table to pick a first domain. Start where typical autonomy is highest and actions are easiest to reverse, which usually means cloud cost or Kubernetes operations, then expand. The last column shows where to start in CloudThinker. Autonomy ranges here are typical for the industry in 2026, not fixed limits. Your own levels depend on how good your observability, IaC and runbooks are.

​
Next steps

Chapter 4: Architecture

See how production agent systems are built from orchestrated specialists.

Chapter 6: Trust and Governance

Set the policies and guardrails that decide which autonomy level each action gets.