Agentic Technologies

“Primates evolve over millions of years, I evolve in seconds... In exactly four minutes… I will be everywhere.” (Terminator Genisys)

The thermostat was probably one of earliest agentic technologies: it has a policy (decision-making rule), a simple utility function (payoff), and very limited mechanism of adaptation (environmental tracking). More advanced agentic technologies — including organisms, entrepreneurs, and reinforcement learning (RL) systems with curiosity-driven exploration more generally – include a mechanism of invention not just a preordained policy. Any advanced agent has the character that the philosopher Emmanuel Kant called “Naturzweck”, natural purposiveness: the system is both cause and effect of its own organization.

All agents contain at least three coupled components: (1) a policy space over which it can search, (2) an objective or utility function that evaluates its performance in this space, and (3) a means of generating new candidate policies. The third element, an open-ended ability to generate novel policies is what distinguishes a genuinely agentic generative system from a simple reactive system or even one capable of limited adaptation.

When we speak of agentic technologies we have in mind creative learning systems that operates and execute persistently and autonomously in changing environments and that are capable of sharing insights with other agents.

And advanced Agentic technologies force us to consider the question of “free will”: are agents free to the degree that their actions are generated by internal models rather than directly caused by external forces. And if so does human freedom and technological freedom align?