AI Agent Taxonomy: How should we classify AI Agents?
Agents, agentification, agentic AI—and the one I prefer, the agentverse—bundle disparate ideas in an attempt to simplify but often obscure detail to the point that the term loses meaning. I developed this AI agent taxonomy to offer clarity, with caveats about overlap and conceptual conflation meant to reflect the complexity of this emerging market.
Most real-world agents could appear in multiple categories. For instance, a Customer Agent might be Collaborative, Functional, Human-Augmented, and Learning. Agent implementations would then be better represented by a multi-dimensional matrix, with the taxonomy framing the matrix.
How Agents Reason and Respond
Reactive Agents operate based on the “condition-action” principle, reacting to immediate inputs and their current perceptions according to predefined rules. They do not maintain a deep understanding of the world beyond the current situation and typically do not consider past actions or future consequences. Fixed automation agents that execute pre-programmed instructions with predictable behavior and limited scope fall under this category. Simple reflex agents are a prime example.
Deliberative Agents engage in reasoning and planning to achieve specific goals. They often maintain an internal model of their environment to evaluate potential actions and their consequences before acting (model-based reflex agents). Goal-based agents fall under this class, as they compare different approaches to choose the most efficient path to a desired outcome.
Utility-based agents, which aim to maximize a predefined utility function by balancing competing goals, are also included here. The Salesforce Atlas reasoning engine, which combines multiple LLMs and LAMs to break down high-level questions into steps and execute them, should be considered a platform for deploying deliberative agents. Planning paradigms like ReWOO, where agents anticipate steps and collect tool outputs before formulating a response, also fit this category.
Agent Capabilities and Outputs
Generative Agents primarily focus on creating new content or solutions autonomously leveraging advanced machine learning models like GANs and transformer models. Types include text-generative agents (e.g., GPT-4), image-generative agents (using GANs), audio-generative agents, and decision-making agents that provide strategic plans. Key characteristics include creativity, adaptability, autonomy, and context awareness in generating novel outputs. Some might argue that “generation” is a modality, not a type of agent, but given the focus on outcomes, I will leave this classification in place to facilitate discussions about agent applications that use generative AI as a starting point.
Learning and Adaptive Agents can learn from data, interactions, and feedback and improve their performance over time. They can refine their responses, adapt to new scenarios, and continuously enhance their capabilities. Learning can occur during development (e.g., fine-tuned models) or during operation (e.g., reinforcement learning or continual learning). The concept of iterative refinement through feedback mechanisms is central to this class.
The Serious Insights AI Agent Taxonomy
AI Agent Taxonomy: Use-Case Archetypes
These agents are categorized by their specific business role or the type of tasks they are designed to perform within an organization.
This includes agents such as the ones found in these three categories:
Information Agents
- Document agents (extracting and organizing data from documents)
- Database agents (turning raw data into insights)
- Report agents (generating content)
- Unstructured agents (processing unstructured data)
Interaction Agents
- Conversational agents (providing client and employee support)
- Customer agents (streamlining customer support)
Operational Agents
- Decision agents (making business decisions)
- Data agents (combining data and tools for insights)
- Code agents (assisting with software development)
These agents often combine reasoning, planning, and learning to fulfill their specific functions. To be clear, this list of generalized agents is not a technical classification but a precursor to business-oriented classification for agent implementations that would be represented by use-case profiles and vertical applications.
Operational Structures and Interaction Modes
Collaborative and human-augmented agents aren’t standalone categories—they are control modes or operational overlays that can apply to agents of any type.
For example, a learning agent that incorporates human feedback becomes both learning and human-augmented by design.
Collaborative Agents (Multi-Agent Systems): These agents operate within systems comprising multiple AI agents interacting to achieve a common objective or individual goals. They can collaborate, communicate, and even compete. Hierarchical agents, where higher-level agents direct lower-level agents, and multi-agent frameworks like Swarm, CrewAI, and AutoGen are key examples.
These systems leverage the diverse capabilities and roles of individual agents to tackle complex tasks and can simulate human behaviors like interpersonal communication. The integration of various AI Agents like Document AI, Decision AI, Database AI, Conversational AI, and Unstructured AI to streamline enterprise operations also exemplifies this class.
Human-Augmented Agents: These agents are designed to interact with humans, incorporating human feedback, guidance, or oversight at specific points in their workflow. This “human-in-the-loop” (HITL) approach is crucial for ensuring ethical operation and alignment with organizational values, especially in sensitive tasks. Frameworks like LangGraph support seamless human intervention. These agents balance autonomy with necessary human control. Interactive partners or surface agents that assist users with tasks and are often triggered by user queries also fall under this category.
AI agent taxonomy types from other classification systems
Levels of AI Agents (L0-L5): A level-based system (L0 – No AI with tools, L1 – Rule-based, L2 – IL/RL based with reasoning, L3 – LLM-based with memory, L4 – Autonomous learning, L5 – Personality and collaboration) is a granular framework based on the evolution of AI core capabilities. It’s less about distinct categories and more about a developmental spectrum.
Agents based on Architectural Paradigms (e.g., ReAct, ReWOO): These are reasoning and action workflows or paradigms used within the architecture of various agent types, particularly deliberative and collaborative agents. They describe how an agent reasons and acts rather than defining a distinct class of agent based on its fundamental characteristics or purpose. We choose to place agents implemented from those architectures into the corresponding functional class.
Agents defined solely by the framework used to build them (e.g., LangChain Agents): While frameworks like LangChain, CrewAI, and Autogen provide the tools and structures for building various types of agents, the framework itself doesn’t define a fundamental class of agent. An agent built with LangChain could be reactive, deliberative, functional, or collaborative.
Agents defined by specific technical implementations (e.g., agents using specific LLMs or tools): Similarly, agents might be described by the specific LLMs they use (e.g., LLM-based agents) or the tools they integrate with. This is a characteristic of an agent, not a primary classification in itself. It is increasingly likely that agent platforms will be LLM-agnostic, making the framework (above) more distinctive than the LLM or suite of LLMs called during agent execution.
This taxonomy is not exhaustive or prescriptive. It’s a functional lens for understanding an evolving market—intended to support design, selection, and strategic discussion.
For more serious insights on AI, click here.
Cover image by ChatGPT from a Daniel W. Rasmus prompt.
Leave a Reply