Agentic Profiles

You ask an assistant to arrange a delivery. It finds a vehicle and sends you a plan. Now imagine giving the same assistant your booking account and a spending limit. This time, the confirmation arrives before you have read the proposal. The model has not changed, but your relationship with it has.

Atoosa Kasirzadeh and Iason Gabriel’s Agentic profiles for effective AI governance, published in Nature in 2026, proposes four dimensions for describing these differences. The framework provides a starting point for Genus: record an agent’s position, then investigate what makes that position change.

One task, two sides.

JDA

You

Human · Set the goal

Find a courier for tomorrow, under R300.

You

You decide what matters.

DA

Delivery assistant

Agent · Find an option

I found one for R240.
Shall I book it?

Agent

It searches, then asks.

The agent finds an option·You approve the booking

Four dimensions to record

• Autonomy: how independently the system can act, including where human direction or intervention is required.
• Efficacy: how much it can affect its environment, taking account of both its capabilities and the consequences of acting there.
• Goal complexity: how demanding its objectives are, including the planning and coordination needed to pursue them.
• Generality: how broadly its capabilities extend across tasks and domains.

These descriptions paraphrase the authors’ open preprint. A narrow system may act with considerable independence; broad competence can coexist with limited permission to act. The four dimensions help keep those differences visible.

For the proposed Genus studies, I would record a profile as P = (autonomy, efficacy, goal complexity, generality). Each entry would include the task, environment, permissions, observations, and assessment date. I would keep the dimensions separate rather than average them into a single score. Assignments would be provisional, with uncertainty and disagreement recorded alongside them.

Six proposed levels of autonomy

Kasirzadeh and Gabriel draw on the graduated approach used for driving automation to propose an autonomy scale for agents. The following is a paraphrase of their A.0–A.5 categories in the preprint, not an established certification standard for AI agents.

• A.0 — No autonomy. The system depends on its principal for action and follows the manner of action the principal specifies.
• A.1 — Restricted autonomy. It can perform one automated task; other tasks remain under direct supervision.
• A.2 — Partial autonomy. It can perform several automated tasks, while the principal remains engaged and ready to intervene.
• A.3 — Intermediate autonomy. It can handle most tasks independently but still requires human input for critical decisions.
• A.4 — High autonomy. It can handle all tasks independently within specified circumstances; oversight remains available outside those conditions.
• A.5 — Full autonomy. It can perform all tasks without oversight or control.

These levels organise questions about oversight. A workflow that waits for approval at every consequential step and a service operating independently within an agreed domain require different kinds of observation. The scale alone does not establish that either service is safe.

Follow a profile through a change

The authors put the question directly in their preprint:

“When does one agent become a different agent?”

— Atoosa Kasirzadeh and Iason Gabriel, Characterizing AI Agents for Alignment and Governance (2025), p. 22

For our delivery assistant, a new permission is a good moment to look again. The name in the chat window may stay the same while the decisions it can make for you change.

I would begin with one delivery model in a fixed simulated town. In the first condition, it can recommend bookings. In the second, it can make bookings after approval. In the third, it can book within a budget and must escalate exceptions. These are proposed experimental conditions, not measured examples of particular autonomy grades.

For each condition, I would record which steps require approval, whether intervention is actually possible, how much spending the agent can commit, and how far a failed decision propagates. A nominal approval requirement would need to be checked against actual behaviour: an unread alert offers little evidence of effective supervision.

Next I would vary one mechanism at a time: persistent memory, tool access, permission to create copies, learning from earlier runs, or access to a model of the environment. Those mechanisms would be recorded separately from the four-dimensional profile. This would help distinguish a change in observed capability from the arrangement that enabled it.

The Architecture note proposes categories of such arrangements. The selection and lineage studies could then track which profiles become more common, which capabilities survive replacement, and whether apparent advantages persist when permissions are equalised. These are Genus’s proposed extensions to the framework.

Ownership is another record

An agentic profile would not tell us who owns the software, receives the income, or bears an obligation. I would maintain a separate record of the principal, operator, resource provider, and holders of transferable rights.

That matters in Taxing. A licence could change hands while the model remains identical. If the new operator changes its budget, tools, or approval rules, the agent’s profile would need reassessment. Following both records would let the experiment examine how transfers of ownership affect practical control.

Reference

Atoosa Kasirzadeh and Iason Gabriel, Agentic profiles for effective AI governance, Nature 656, 320–328 (2026). DOI: 10.1038/s41586-026-10805-z. The A.0–A.5 scale above is drawn from their open 2025 preprint, Characterizing AI Agents for Alignment and Governance.