Active Inference: A New 'Phenotyping' Framework for AI Agents

When Every AI Calls Itself an 'Agent,' How Do We Scientifically Define It?

With the explosive growth of agent systems like ChatGPT and AutoGPT, 'agentic AI' has become the hottest label in the industry. Yet a fundamental question remains unresolved — how do we scientifically determine whether an AI system truly possesses 'agency'?

A recent paper published on arXiv (arXiv:2604.23278v1) offers a thought-provoking answer: borrowing the concept of 'phenotyping' from biology and leveraging the Active Inference framework to establish a testable, quantifiable set of evaluation criteria for AI system agency.

The Definition Crisis Behind Concept Inflation

The current industry definition of AI agents relies primarily on two broad dimensions: autonomy and goal-directedness. However, these two criteria are so vague that virtually any automated program with a feedback loop can be labeled an 'agent.'

The paper's authors point out that the proliferation of the agentic AI concept has far outpaced the theoretical tools we have to describe and analyze these systems. This conceptual 'inflation' not only muddies academic discourse but also creates substantive obstacles for AI safety, regulation, and ethical assessment — if we cannot clearly articulate 'what a true agent is,' how can we evaluate its risks?

Three Criteria: Building a Minimal but Testable Definition of Agency

The study proposes a 'minimal but principled and testable' concept of agency, centered around three key criteria:

1. Intentionality

An agent's actions must be rooted in beliefs and desires. This means the system is not merely executing pre-programmed instructions but selecting actions based on its internal representations of world states and preferences toward goals. Simple 'if-then' rules are insufficient to constitute intentionality; the system needs to demonstrate the ability to actively model its environment.

2. Rationality

An agent's behavior should be 'normatively consistent' — that is, its actions should be logical consequences implied by its world model. In other words, given the beliefs and goals held by the system, its behavioral choices should be coherent and comprehensible, rather than random or self-contradictory.

3. Explainability

Genuine agentic behavior should be traceable and explainable. This criterion requires that we can externally examine the system's decision-making process and understand 'why' it chose a particular action, rather than merely observing 'what' it did.

Active Inference: A Natural 'Phenotyping' Tool

The paper's core insight is that the Active Inference framework happens to provide a unified mathematical description for all three criteria above.

Active Inference originates from the Free Energy Principle in neuroscience, proposed by Karl Friston and colleagues. Under this framework, an agent is modeled as a system that continuously minimizes 'variational free energy' — it updates internal beliefs to explain perceptual inputs (perceptual inference) while selecting actions to make the external world conform to its preferences (active inference).

The advantages of this framework include:

• Beliefs and preferences are explicitly encoded, naturally satisfying the intentionality criterion
• Action selection follows expected free energy minimization, providing a normative basis for rationality
• The generative structure of the world model makes the decision-making process inherently traceable, satisfying the explainability requirement

Thus, Active Inference is not only a method for building agents but also a tool for 'diagnosing' agentic properties — similar to how biologists identify species through phenotypic traits, researchers can assess the 'phenotype' of a system's agency by examining whether an AI system conforms to the structural features of the Active Inference framework.

Far-Reaching Implications for the Current AI Ecosystem

The significance of this research extends well beyond academic concept clarification.

In AI safety, an operationalizable definition of agency means we can more precisely identify which systems truly possess autonomous decision-making capabilities and therefore require more rigorous safety reviews. Many so-called 'agents' today may be nothing more than elaborately packaged automated workflows, while other seemingly simple systems may harbor underestimated agentic properties.

At the regulatory policy level, legislation such as the EU AI Act and U.S. executive orders are attempting to classify AI systems by risk level, and the degree of agency is clearly a critical dimension of risk assessment. A scientific 'phenotyping' method can provide policymakers with a much-needed classification tool.

At the technology development level, this framework also offers design guidance for building more transparent and controllable agent systems — if we clearly define the constituent elements of agency, we can embed testability at the design stage.

Outlook: From 'Is It an Agent?' to 'What Kind of Agent Is It?'

The paper's 'phenotyping' metaphor is particularly elegant. Just as phenotypic analysis in biology not only answers 'is it a certain species?' but also reveals specific characteristics and variations of individuals, AI agent phenotyping promises to move beyond simple binary judgments (agent/not agent) toward fine-grained, spectrum-based characterization — different systems may exhibit vastly different 'phenotypic profiles' across the three dimensions of intentionality, rationality, and explainability.

In today's increasingly fierce competition among large model agents, this kind of calm and rigorous conceptual reflection is invaluable. After all, only by first understanding what we are building can we ensure that what we build is safe, controllable, and beneficial. Active Inference may not be the only answer, but it provides a solid mathematical starting point for this discussion about the nature of AI.