Where an Agent’s Power Comes From

Give two agents the same model and ask them to arrange a delivery. One can produce a plan in a text window. The other can check stock, book a vehicle, remember earlier delays, and authorise payment. Their practical power differs before either begins to reason about the route.

Genus starts with those differences. I want to compare the arrangements that let an agent act, learn, and retain an advantage. A classification should earn its place by explaining concrete cases.

The Agentic Profiles note introduces the four dimensions I would track through these comparisons: autonomy, efficacy, goal complexity, and generality. It also explains the proposed A.0–A.5 autonomy scale. The categories below describe arrangements that might change a profile; they are separate from those autonomy levels.

The following four groups are provisional research categories. They may overlap: an agent can have direct tools, learn from observation, and be copied into a larger population.

1. Genus A — Agents with direct tools. These can act on external systems through browsers, code execution, service interfaces, or payment permissions. The relevant comparison holds the model constant and changes what it can reach. A delivery agent with booking access could be compared with one that must ask a person to carry out every booking.

2. Genus B — Composable agents. These are simple units that can be combined or copied into a larger process. Separate units might check stock, compare routes, and inspect invoices. I would compare that arrangement with one more capable agent, including the cost of passing information between units.

3. Genus C — Adaptive populations. These permit configurations to be modified, copied, and retained across rounds. A successful delivery strategy might spread because more agents are instantiated with it. The selection stream examines how resource limits and copying rules determine which strategies persist.

4. Genus D — Observational learners. These learn from records of other agents’ actions and outcomes. An observer might study failed deliveries and produce a reusable account of where plans went wrong. Its value would depend on whether another agent can use that account to improve a later decision.

A shared delivery world would let these groups encounter one another. Tool access might create a lasting advantage, or a large group of inexpensive agents might compensate through coordination. Observers might become useful keepers of experience for agents that are frequently replaced.

I would compare completed deliveries, resource costs, dependence on intermediaries, and the persistence of advantages after tools or records are shared. The purpose is to identify which differences matter, and which disappear when access changes.