Four AI autonomy levels—Guarded, Bounded, Delegated, Utility—and guidance on applying them.
Autonomy isn't a switch, it's a gradient. Four levels from holding the AI's hand to letting it run as infrastructure.
What It Is
How much can the AI decide on its own, without you standing next to it? Simon Wardley sorts this into four levels (2025), running from a tight grip to almost invisible infrastructure.
- Guarded: You hold the AI by the arm. Every suggestion gets checked before execution. Tab-completion in Cursor or Copilot is the classic example
- Bounded: The AI acts on its own within clear limits. "Only change files in src/, ask before commits." You set the guardrails, watch, step in when the boundary breaks
- Delegated: The AI runs independently, escalates only on uncertainty or pre-defined trigger points. You're not in the workflow, you're on call
- Utility: The AI runs as infrastructure. Nobody thinks about it actively anymore. Spell-check, smart-compose, inbox sorting, AI disappears into daily life
The point: control isn't given up, it's transformed. At every level, it's distributed differently, but it doesn't vanish. Reading the model as "eventually drop it entirely" builds Agent Sprawl on purpose.
The level is a property of the task and the current trust, not the person. The same individual might be in Guarded for production code and in Delegated for a research task. Locking levels by employee misses the pattern.
How To Spot It
- The org talks deliberately about which AI sits on which level in which context
- Levels shift as trust grows, without "the AI is allowed more now" drama
- Different tasks visibly run in different levels, no "AI is the same everywhere"
- Escalation paths and telemetry make it visible where Delegated/Utility actually holds up
What To Do (FL2 - Coordination)
- Pick the level per task type: Which workflows need Guarded, which can run Delegated? That's workflow architecture, not gut feel
- Align to the frontier: Deep inside the Jagged Frontier Delegated/Utility is fine, near the edge Guarded is required
- Guardrails before delegation: Without Guardian Agents, Delegated is naive. Safety patterns first, then unlock more autonomy
- Treat level changes as transitions: Lifting an AI from Bounded to Delegated needs adjustments in monitoring, escalation, accountability, not a simple toggle
The Trap
Reading the levels as a maturity model, "higher = better". Utility in the wrong place can be catastrophic. An org pushing everything toward Utility is optimizing for efficiency and missing the risk. The right level follows the task's risk profile, not the roadmap's ambition.