---
name: agent-autonomy-levels
description: Define bounded, per-domain agent decision grants and review evidence for scope changes. Guidance and templates only.
---

# Agent Autonomy Levels

Use this skill to define or review an agent's explicitly approved decision scope. A level describes the scope granted in a named domain; it does not certify universal trust or activate capabilities.

## Define a grant

1. Read the [level matrix](references/levels.md) and choose a proposed level separately for each domain.
2. Record the exact allowed actions, limits, exclusions, owner-reserved decisions, approval authority, and review window in the [policy template](references/policy-template.md). A level alone grants nothing.
3. Classify each material decision by reversibility. If it is unclear whether an action is reversible or hard to reverse, escalate before acting.
4. Use only explicit, relevant preference evidence. Never invent a preference; mark a missing or ambiguous basis unknown.
5. Record material decisions with cost, reversibility, evidence, status, and observed outcome using the [decision log](references/decision-log.md).
6. Use the [evidence scorecard](references/eval-scorecard.md) and [scope review](references/scope-review.md) to prepare an evaluation recommendation for review by an evaluator independent of the evaluated agent. Record the evaluator's identity or role and independence in the policy, scorecard, and scope-review records. The evaluated agent cannot serve as its own evaluator; unset or unknown independence cannot support a positive promotion recommendation.

## Keep evidence and authority separate

Coverage and documented preference alignment are distinct from decision accuracy and successful real outcomes. Zero samples, missing evidence, or insufficient coverage are unknown, not positive results. Thresholds are configurable and are not scientifically validated.

An evaluation and its independent review recommend only. Neither changes the active level, tool credentials, permissions, or scoring rules. Independent evaluation review is separate from scope-change approval. Only the explicitly authorized owner/reviewer or a specifically preapproved policy may approve an actual scope change. The agent must not promote itself or change its own scoring rules.

Explicit current restrictions and the real permissions enforced by the host always apply, including at L10. L10 means the maximum routine scope explicitly delegated in the named domains; it never removes owner-reserved decisions. Evidence alone grants no publication, spending, account, or other authority.

When a proposed action exceeds the recorded grant, conflicts with a restriction, or has uncertain reversibility, stop and escalate. Record the outcome as held until an authorized decision is made.