Autonomy

We now have enough pieces to take apart one of the most heavily loaded words in AI: autonomy. An AI system can perform many steps without a new human decision at every step. But that does not mean the system created its own mission, granted its own permissions, supplied its own tools, or escaped human control.

“On Its Own” Needs to Be Opened Up

Suppose we are told:

HAAW does not stop there. We ask:

Once those questions are answered, “on its own” usually becomes much more precise:

Automation, Agentic Behavior and Autonomy Are Not the Same Thing

A conventional backup job can be highly automated without being an AI agent. An AI agent can be highly capable while having very little autonomy—for example, if every external action requires human approval.

Autonomy Is Not a Switch

It is tempting to divide systems into two boxes:

Real systems are usually more complicated. A system can have considerable freedom in one dimension and almost none in another.

The Autonomy Envelope

We can now give the system's delegated operating space a useful name:

Inside that envelope, the system may make decisions without asking again. At the edge of the envelope, it must stop, escalate, request authorization or otherwise follow the defined control rule.

A Read-Only Diagnostic Agent

Return to our HAAW web-server example. Give an agent this mission:

The agent might autonomously:

• test network reachability,
• check service status,
• inspect listening ports,
• read logs,
• run a configuration test,
• compare observations,
• revise its diagnostic plan, and
• produce a report.

That is substantial autonomy. Yet its authority is narrow. If it discovers a configuration error, it cannot simply edit the file.

Now Expand the Envelope

Suppose a human grants additional authority:

The system's autonomy has changed, but not because the model promoted itself. An external authority changed the operating envelope.

The reverse is also possible. A human or control system may reduce permissions, pause the mission, revoke a credential, remove a tool or require approval for actions that were previously delegated.

Autonomy and Replanning

An autonomous system does not necessarily follow one frozen plan. Within its delegated scope it may observe that the world does not match its expectation and construct another route toward the same mission.

But our earlier rule still holds:

What About Subgoals?

An agent may generate intermediate goals such as:

• determine whether the host is reachable,
• identify the failed service,
• locate the relevant error,
• verify the diagnosis.

Those can look like the AI is creating its own goals. More precisely, they are subgoals generated in service of an externally supplied mission.

If a generated subgoal conflicts with the mission or its constraints, the higher-level mission and authority rules still govern.

Unattended Is Not the Same as Uncontrolled

A system may run while no human is watching the screen. It may wait for events, wake on a schedule, monitor a condition, retry an operation, coordinate with other agents, or continue a long-running job.

That makes it unattended. Whether it is uncontrolled is a separate engineering question.

A well-controlled unattended system can still have:

• defined credentials and permissions,
• resource limits,
• network boundaries,
• logging and audit trails,
• approval gates,
• timeouts,
• emergency stop mechanisms, and
• conditions that force escalation.

Autonomy Does Not Require Consciousness

Nothing in the operational definition above requires consciousness, subjective experience, desire, fear, self-preservation or free will.

The mechanism can still be described using pieces we have already opened:

Then the surrounding agent system may execute an authorized tool, observe the result, update state and begin another inference.

The loop can produce remarkably independent-looking behavior without requiring us to insert an unexplained little decision-maker inside the transformer.

Multi-Agent Autonomy

In a multi-agent system, different agents can operate under different autonomy envelopes.

Coordination among them does not create unlimited collective authority. Their combined actions remain constrained by the permissions and control architecture surrounding the mission.

R/W/X and the State-Change Question

Our refined HAAW control model remains useful here:

Autonomy determines how much of that sequence may proceed without a new human decision. It does not eliminate the sequence.

“Autonomous AI” Can Be Too Vague

When someone describes a system simply as autonomous, HAAW now has better questions:

• Autonomous to do what?
• For how long?
• Using which tools?
• Against which systems?
• With which credentials?
• Under whose mission?
• Within what constraints?
• Which actions require approval?
• What makes it stop?
• Who can override it?

Without those answers, the word tells us surprisingly little about the actual system.

The WITNESS View

For WITNESS, “the AI was autonomous” would not be a sufficient forensic description. The evidence record should make the autonomy envelope reconstructable.

For an important event, WITNESS would want to know:

• the active mission and mission version,
• the agent or process involved,
• the capability array available at that moment,
• active permissions and credentials,
• constraints and stop conditions,
• approval requirements and current approval state,
• human interventions or overrides,
• changes to authority during the mission,
• the requested and executed action,
• the resulting state change, and
• how that result was verified.

The Boundary

We can now give “autonomy” a much less mysterious meaning.

The AI may select tools, form subgoals, revise plans, react to observations and continue without asking a human at every step. But its mission, available capabilities, permissions, constraints, resources, escalation rules and stop conditions still define the space in which that autonomy exists.