Back to Blog
Agentic AI4 min read
When Someone Says They Deployed an "Autonomous Agent," Ask One Question
August 15, 2026
When someone tells you they deployed an "autonomous agent," ask one question: at what level of autonomy?
Autonomy is not an on/off switch. It is a scale.
Cloud Security Alliance's autonomy framework lays out a six-level taxonomy adapted from autonomous vehicles:
Human oversightAgent autonomy
L5Full Autonomy
Agent sets its own goals
L4High Autonomy
Human monitors only
L3Conditional
Agent acts, escalates at boundaries
L2Supervised
Human approves the plan
L1Assisted
Human approves each action
L0No Autonomy
Human execution
L0No autonomy
The agent informs; the human executes every single action. Think standard search and informational chatbots.
L1Assisted
The agent performs distinct tasks, but every individual step requires separate human sign-off.
L2Supervised
The human approves the overall plan upfront; the agent executes independently within strict boundaries. This is the dominant enterprise pattern today.
L3Conditional
The agent operates autonomously within guardrails and escalates only on exceptions. Oversight shifts from approving every step to handling edge cases.
L4High autonomy
The agent executes across broad workflows; humans review audit logs retrospectively rather than approving actions upfront.
L5Full autonomy
The agent defines its own goals and sub-tasks without operational boundaries. Human oversight is purely strategic.
Nobody runs L5
The reality across enterprise production environments: nobody runs L5. The framework itself notes L5 is not suitable for enterprise deployments. Most enterprises sit at L1-L2 today. The incidents that make headlines happen when agents get pushed to L3-L5 without controls to match.
The operational rule
Never set autonomy at the agent level. Set it per action.
Base it on reversibility, blast radius, and whether a human can catch a failure before it impacts data or capital.
Generating an internal summary? L3 works. Triggering database migrations, transferring funds, or emailing clients? Keep it at L1 or L2 with a hard approval gate.
Autonomy is an operational governance decision, not a tech spec.