A Spectrum, Not a Switch: Five Levels of Agency

2026-05-07
Section 1 · TaxonomySlide 4 of 22

Agency is a spectrum, not a switch.

Borrow the SAE driving-autonomy framework. Five levels, each with a clean answer to one question: who decides, and what is the human's role?

LevelL0

Suggestion

What the AI decides

Nothing. Generates text or completions a human reads.

Illustrative example

Inline code completion. Draft email suggestions.

LevelL1

Single-Step Tool Use

What the AI decides

Which one tool to call, with which arguments, in response to a prompt.

Illustrative example

“Look up this customer's order status.” One API call, one answer.

LevelL2

Multi-Step Workflow

What the AI decides

A planned sequence of tool calls within a bounded scope, with human approval at the end.

Illustrative example

“Draft a refund. Show me the proposal.” Plans, gathers data, waits for sign-off.

LevelL3

Bounded Autonomy

What the AI decides

When to act, which steps to take, when to retry — within a guardrailed budget. Human is on-the-loop.

Illustrative example

An incident-response agent that triages alerts and opens tickets, escalating when confidence drops.

LevelL4

Goal-Directed Autonomy

What the AI decides

Goal, plan, sub-goals, and self-correction. Human reviews outcomes, not steps.

Illustrative example

A research agent given a topic and a budget. Returns a deliverable hours later.

Practical reality: most enterprise value in 2026 lives at L2 and L3. L4 looks impressive in demos and fragile in production.