Agentic AI

What makes an artificial intelligence an agent rather than simply an intelligent tool?

AI systems can answer questions, write code, make plans and use tools. But intelligence and agency are not the same thing. A system may be extraordinarily intelligent while having little ability to shape what happens next.

Agentic AI: A Possibility-Based Theory of AI Agency explores a different way of thinking about artificial agency: through the futures an AI system can actually reach, preserve, create and change.

Using the simple example of a town preparing for an uncertain flood, David Glacier develops the idea of possibility-structured agency. An agent does more than choose the best option from a fixed menu. It may need to recognize which futures are genuinely reachable, preserve valuable capabilities, create new routes, investigate uncertainty, adapt when circumstances change-and know when it is finally time to commit.

But does the idea survive testing?

The book follows a series of reproducible computer experiments designed to put parts of the theory under pressure. Thousands of artificial decision situations explore what happens when priorities change, information has a price, learned judgement encounters unfamiliar conditions, recovery becomes difficult and important possibilities can be lost forever.

The results reveal an important complication: there is no universally best strategy.

Keeping more options open can help-or become expensive and wasteful. Gathering more information can improve decisions-or arrive too late to matter. Adaptation can outperform careful initial planning, while uncertainty can justify caution only when the consequences warrant its cost.

The deepest question is therefore not simply whether an AI can achieve a goal. It is whether an increasingly capable system can manage the changing landscape of possibilities surrounding that goal while remaining bounded by human authority, safety and real-world constraints.

Written for the general reader, Agentic AI develops the concepts first and places the mathematics and experimental machinery in separate technical chapters for readers who want to examine them.

This is not a claim that the mathematics has solved agency, nor that simplified simulations prove how real AI systems will behave. It is the beginning of a testable framework-and an invitation to ask a different question about the future of artificial intelligence:

Not merely, "What can this AI do?" but "What futures can it help keep possible?"