// FIELD NOTE / 01

Autonomy is an authority decision.

Last updated: August 23, 2026

An agent does not become operationally ready because a model can complete a task. The deployment decision is about delegated authority: what the agent may observe, decide, change, spend, disclose, or trigger - and who remains accountable when conditions move outside the expected envelope.

Start with the minimum useful authority

Use the lowest level that produces the mission outcome. A human-assisted workflow may be sufficient for high-consequence work. Approval-gated or semi-autonomous operation can shorten cycle time without granting broad execution rights. Bounded autonomy is appropriate only when the workflow is narrow, observable, reversible, and supported by tested controls.

1. Human-assisted: the system advises; a person decides and acts.
2. Human-in-the-loop: the system prepares an action and stops at defined approval gates.
3. Semi-autonomous: the system executes within policy, tool, data, and budget limits; exceptions route to a person.
4. Bounded autonomous: the system runs a narrowly defined, reversible software workflow with continuous evidence and a fail-safe path.

The authority baseline

Before deployment, record the intended outcome, accountable owner, permitted subjects and objects, allowed tools and data, prohibited actions, approval thresholds, budget and time limits, monitoring signals, escalation conditions, rollback behavior, and evidence-retention requirements. Acceptance should attach to a signature-ready, agency-owned baseline—not to a vague promise of responsible AI.

More autonomy is not automatically more mature. The mature choice is the least authority that reliably delivers the mission result.

Federal policy context

OMB Memoranda M-25-21 and M-25-22 establish current federal AI-use and acquisition context, including risk management, performance, competition, and cross-functional acquisition. They do not replace the system-specific authority decision for an agent acting through enterprise tools. NIST's AI Agent Standards Initiative further emphasizes secure, interoperable agents, identity, authorization, and evaluation.

Practical output

The useful artifact is a decision-ready authority baseline with testable acceptance criteria, not another slide of principles. Approval and risk acceptance remain agency decisions. The baseline should be revisited whenever the agent receives a new model, tool, data source, permission, workflow, environment, or operating objective.