Autonomy Levels

Posture: docs_first, advisory_first. This document describes vocabulary and required behavior. It does not implement a runtime, a permission engine, or any tool. It is provider-agnostic.

Goal

Name the levels of authority an agent may hold, so a skill can declare a ceiling and a harness can enforce one. Autonomy is about authority, not depth. How deep a task runs (effort) is governed by the Runtime Effort Governance Contract; how much authority the agent holds is this axis.

Canonical levels

These four values are the source of truth and are enforced structurally by the Agent Harness Contract schema (enum: ["observe", "advise", "act_with_approval", "autonomous_within_bounds"]) and described in agent-harness-contract.md and ADR-013. Do not invent additional level names.

autonomy_levels:
  observe:
    description: "Can inspect context and summarize. No recommendations that trigger action."
  advise:
    description: "Can recommend options and record decisions. No tool side effects."
  act_with_approval:
    description: "Can prepare or execute bounded actions only after explicit approval gates."
  autonomous_within_bounds:
    description: "Can act within narrow, pre-approved constraints; must declare blocked actions and rollback."

The schema enforces that autonomous_within_bounds declares at least one blocked action and requires rollback (Invariant 3). Treat unbounded autonomy as out of scope for this phase.

Vocabulary reconciliation

External material (the Agent Harness Enforcement Addendum drafts) sometimes uses a five-level model with bounded_autonomous and a standalone autonomous. Map them to the canonical vocabulary:

Skills in the Adaptative Skills repo declare an autonomy floor and ceiling using these canonical names in their per-skill harness_requirements (see Adaptative Skills docs/harness-requirements-for-skills.md).

See also

• ADR-014 — the reconciliation decision recorded.
• agent-harness-contract.md — per-task envelope that carries the level.
• tool-risk-taxonomy.md — the orthogonal “how risky is this tool” axis.
• enforcement-boundaries.md — who declares vs. who enforces.