Agentic Autonomy Levels is a six-level scheme for classifying how autonomously an AI coding agent (or fleet of agents) operates, built from two axes measured separately: agency, how far a single agent is allowed to go before a human decision is required, and orchestration, how many agents run and who coordinates them. The two axes are read together as one climb because, in this scheme, orchestration only becomes a distinguishing factor near the top of the scale.
Usage
The six levels are: Level 0 (Assist) — the agent suggests actions and a human decides on every one; Level 1 (Supervised action) — the agent edits or runs commands but asks before anything consequential; Level 2 (Scoped task delegation) — a bounded task with a clear goal and definition of done is handed off, verified by evidence such as passing tests; Level 3 (Goal-driven autonomy) — the agent does whatever it takes to reach a goal defined by a measurable, automatable stopping condition; Level 4 (Parallel delegation) — multiple agents work isolated slices of a task in parallel; and Level 5 (Managed-by-exception orchestration) — a manager agent dispatches worker agents against defined policies, verifies their output, and escalates only exceptions to a human.
The scheme is presented as a two-axis alternative to Steve Yegge’s earlier single-axis autonomy ladder (from “Welcome to Gas Town”), on the grounds that a single number cannot separately represent an individual agent’s trust level and an organization’s skill at coordinating many agents at once.