In 2001, I submitted my MIS Honours thesis. The title was Programming Intelligent Agents with Assumptions.
A few weeks ago, I was on a podcast explaining to an audience what an AI agent actually is.
Twenty-five years. Same question.
The thesis was about deliberative agents: systems that don’t just react to input but hold beliefs about the world, act on those beliefs, and revise them when new information arrives. Getting that to work reliably, even in a constrained environment, was genuinely difficult. The formal methods – predicate logic, belief revision frameworks – were elegant on paper and humbling in practice. You could build a simulated agent that navigated uncertainty in a controlled toy world. It was not going to survive contact with anything real.
Over the following two decades, the formal logic-based approach I had been trained in was displaced by statistical and neural methods. The new approach did not try to explicitly represent knowledge; it trained on scale and let behaviour emerge. The results were narrow for most of that period: capable at specific tasks, brittle outside them. They could beat the world’s best Go player without being able to explain a single move.
Then, around 2022, something qualitatively different appeared. Language models that could reason across domains, hold context, work with incomplete information – not because they had been programmed to reason, but because at sufficient scale, reasoning emerged anyway. The 2001 problem had been solved, by a method I would not have recognised as a solution at the time.
What has not changed is the set of hard questions underneath it.
How do you know what an agent knows versus what it is confabulating? The modern term is hallucination; the 2001 term was belief revision under uncertainty. How do you bound an agent’s behaviour in high-stakes environments where mistakes carry clinical or governance consequences? How do you design accountability for a system that can act in the world, not just produce text?
These are the questions I worked on in 2001, in a postgraduate laboratory in Wollongong. They are the questions I work on now, as Deputy Chancellor at the University of Wollongong and in boardrooms where the failure modes are no longer academic.
The formalism I learned is largely irrelevant to the systems I work with today. The instinct it built – that agents require careful design, that governance is not an afterthought, that you need to understand what a system can and cannot reliably know – that turns out to be exactly the right preparation.
The questions outlast the methods.