TLDR: A new research paper by Christopher Wissuchek and Patrick Zschech introduces a comprehensive typological framework to classify and understand agentic AI systems. Moving beyond passive tools, agentic AI can reason, adapt, and act autonomously. The framework defines agency through interactivity, autonomy, adaptability, and normativity, and categorizes AI systems across eight dimensions (Knowledge Scope, Perception, Reasoning, Interactivity, Operation, Contextualization, Self-improvement, Normative Alignment) with four levels of sophistication. This typology helps researchers and practitioners assess current AI capabilities, guide future development, and address potential risks, providing a structured lens for the evolving landscape of AI.
Artificial intelligence is rapidly evolving, moving beyond simple tools to become autonomous agents capable of reasoning, adapting, and acting with minimal human help. This significant shift, often called the “agentic turn,” highlights a growing need for a clear way to understand and categorize these advanced AI systems. A new research paper, Exploring Agentic Artificial Intelligence Systems: Towards a Typological Framework, by Christopher Wissuchek and Patrick Zschech, introduces a comprehensive framework to do just that.
Understanding Agentic AI
Traditionally, AI systems were seen as passive tools, performing tasks only when explicitly told. However, modern AI, especially with the rise of large models like Large Language Models (LLMs), can now initiate actions, adapt to changing situations, and pursue complex goals independently. Think of GitHub Copilot, which observes a software engineer and adaptively generates code, versus a simple analytical tool that only provides insights when prompted.
The researchers define “agenticness” not as a simple yes/no, but as a spectrum. They identify four core conditions for an AI system to be considered agentic:
- Interactivity: The ability to engage with and be affected by its environment.
- Autonomy: The capacity to function without direct human control.
- Adaptability: The capability to learn from past interactions and refine its behavior over time.
- Normativity: Its behavior being guided by specific goals or objectives.
A New Framework for Classification
To bring order to this complex and fast-moving field, the paper develops a typology – a structured classification system. Unlike a taxonomy that categorizes existing things, a typology helps conceptualize and theorize about phenomena, including future developments. The framework is designed to be technology-agnostic, meaning it can classify any AI system, regardless of its underlying model.
The typology introduces eight key dimensions, each with four levels ranging from “non-agentic” to “general intelligence” (a speculative future state):
- Knowledge Scope: How much and what kind of information the AI has access to (from narrow to actively exploratory).
- Perception: The AI’s ability to perceive inputs (from none to intuitive multimodal understanding).
- Reasoning: The AI’s capability to process and plan tasks (from one-shot responses to theory-of-mind reasoning).
- Interactivity: How the AI engages with its environment (from passive to dynamically engaging with multiple actors).
- Operation: The mode in which the AI system functions (from on-demand to self-organizing).
- Contextualization: The AI’s ability to integrate and retain context (from stateless to holistic human-like situational awareness).
- Self-improvement: The AI’s ability to learn and adapt (from static to evolutionary self-restructuring).
- Normative Alignment: The AI’s alignment with ethical, social, or procedural norms (from unaware to value-aligned).
Putting the Typology to Use
The paper demonstrates the practical application of this framework by classifying real-world AI systems like OpenAI’s Deep Research, Microsoft’s Copilot Agents, Copilot Chat, and Operator. For instance, Deep Research is classified as a “Research Agent” due to its high cognitive agency (reflective reasoning, externally-informed knowledge) but low environmental agency (passive interactivity, periodic operation).
The researchers also simplify the eight dimensions into two overarching categories:
- Cognitive Agency: How the AI system “thinks” – its internal deliberation, learning, and ethical considerations.
- Environmental Agency: How the AI system acts in and interprets its surroundings – its ability to sense, interact, and operate within its environment.
This reduction helps categorize AI systems into four constructed types: Simple Agents, Research Agents, Task Agents, and Complex Agents, providing a streamlined way to understand their core capabilities.
Also Read:
- Unpacking How AI Thinks: A New Framework for Understanding Reasoning Systems
- Noosemia: Exploring the Human Tendency to Perceive Minds in Generative AI
Implications for the Future of AI
This typological framework offers significant benefits for both researchers and practitioners. For researchers, it provides a consistent way to define and compare AI systems, trace their progression, and study their impact. For practitioners, it acts as a decision-support tool, helping organizations select and integrate agentic AI solutions that align with their strategic goals. It also aids in guiding the development of new AI systems by clarifying which capabilities to prioritize.
Furthermore, by breaking down agentic AI capabilities into measurable dimensions, the framework helps in investigating potential risks and harms associated with increasingly autonomous systems, such as shifts in decision-making authority or ethical dilemmas. This can inform the development of appropriate safeguards and governance mechanisms.
As AI technology continues its rapid advancement, this framework provides a robust and adaptable tool for evaluating and shaping the next generation of agentic AI solutions, ensuring a more informed and responsible approach to their deployment.
