Agent Taxonomy

Overview

Systematic classification of agents is fundamental to understanding their design space. From Russell & Norvig's classic four-type taxonomy to Wooldridge's weak/strong agency, and to new classification dimensions in the LLM era, this article provides a comprehensive survey of agent taxonomy.

Taxonomy Overview

graph TD
    A[Agent Taxonomy] --> B[Classic Taxonomy<br/>Russell & Norvig]
    A --> C[Agency Classification<br/>Wooldridge]
    A --> D[Architecture Classification<br/>Reactive/Deliberative/Hybrid]
    A --> E[LLM-Era Classification<br/>Modern Dimensions]

    B --> B1[Simple Reflex]
    B --> B2[Model-Based Reflex]
    B --> B3[Goal-Based]
    B --> B4[Utility-Based]

    C --> C1[Weak Agency]
    C --> C2[Strong Agency]

    D --> D1[Reactive]
    D --> D2[Deliberative]
    D --> D3[Hybrid]

    E --> E1[Single-Turn Conversational]
    E --> E2[Multi-Turn Interactive]
    E --> E3[Autonomous]
    E --> E4[Multi-Agent]

1. Russell & Norvig's Classic Taxonomy

Russell and Norvig proposed four agent types of increasing complexity in Artificial Intelligence: A Modern Approach.

1.1 Simple Reflex Agent

Selects actions directly based on current percepts without maintaining any internal state.

Decision Rule:

\[ a = \text{rule-match}(\text{percept}) \]

where \(a\) is the selected action and \(\text{percept}\) is the current percept.

Characteristics:

Based on condition-action rules (if-then rules)
No memory; relies solely on current observations
Suitable for fully observable environments
Examples: thermostats, simple chatbot keyword matching

Limitations:

Cannot handle partially observable environments
Cannot perform long-term planning
Number of rules grows exponentially with environmental complexity

1.2 Model-Based Reflex Agent

Maintains an internal model of the environment, capable of handling partially observable environments.

State Update:

\[ s_{t+1} = \text{update}(s_t, a_t, \text{percept}_{t+1}) \]

\[ a_{t+1} = \text{rule-match}(s_{t+1}) \]

Characteristics:

Maintains internal state (world model)
Transition model describes how the world changes
Sensor model describes how to infer state from observations
Example: vehicle tracking in autonomous driving

1.3 Goal-Based Agent

Considers goals in addition to the current state when selecting actions.

Decision Process:

\[ a^* = \arg\max_{a \in A} P(\text{goal} \mid s, a) \]

Characteristics:

Introduces goal representation
Requires search and planning to achieve goals
Can handle complex multi-step tasks
Examples: path planning, STRIPS planner

1.4 Utility-Based Agent

Uses a utility function to quantify how desirable different states are, selecting the action with the highest expected utility.

Decision Process:

\[ a^* = \arg\max_{a \in A} \sum_{s'} P(s' \mid s, a) \cdot U(s') \]

where \(U(s')\) is the utility of state \(s'\) and \(P(s' \mid s, a)\) is the state transition probability.

Characteristics:

Utility functions provide a continuous measure of state preference
Can handle multi-objective conflicts and uncertainty
Theoretically optimal but computationally expensive
Examples: rational decision-makers in economics, AlphaGo's value network

From Goals to Utility

A goal is a binary judgment of "satisfied or not," while utility is a continuous measure of "degree of satisfaction." Utility allows agents to select the optimal solution among multiple feasible options, rather than merely finding any feasible solution.

2. Wooldridge's Agency Classification

Michael Wooldridge distinguished between weak agency and strong agency in An Introduction to MultiAgent Systems.

2.1 Weak Agency

Weak agency requires agents to possess the following basic properties:

Property	Description
Autonomy	Operates without direct external intervention
Social Ability	Can interact with other agents (or humans)
Reactivity	Can perceive the environment and respond to changes in a timely manner
Pro-activeness	Can take initiative to achieve goals

2.2 Strong Agency

In addition to weak agency, strong agency includes mental state properties:

Property	Description
Belief	Knowledge and assumptions about the world state
Desire	Set of goals the agent wishes to achieve
Intention	Committed action plans
Emotion	Emotional states that influence decision-making

BDI Connection

Strong agency directly corresponds to the BDI model. See BDI Model for details.

3. Architecture Classification: Reactive vs. Deliberative vs. Hybrid

3.1 Reactive Agent

Does not maintain a world model; maps directly from percepts to actions.

Representatives:

Brooks' Subsumption Architecture
Behavior Trees

Pros: Fast response, simple implementation, strong robustness Cons: Cannot perform long-term planning, difficult to handle complex goals

3.2 Deliberative Agent

Maintains an explicit symbolic model of the world; selects actions through reasoning and planning.

Representatives:

STRIPS/PDDL planners
BDI systems (PRS, Jason)

Pros: Handles complex goals, strong interpretability Cons: Slow response, world model may be inaccurate

3.3 Hybrid Agent

A layered architecture combining a reactive layer and a deliberative layer.

Classic Architecture:

┌─────────────────────────────┐
│  Deliberative Layer (Planning)│  ← Long-term planning, symbolic reasoning
├─────────────────────────────┤
│  Sequencing Layer            │  ← Short-term decisions, coordination
├─────────────────────────────┤
│  Reactive Layer              │  ← Immediate response, emergency behaviors
└─────────────────────────────┘
        ↕ Environment

Representatives:

InteRRaP three-layer architecture
TouringMachines
Modern LLM agents (prompt parsing + tool calling + reflection)

4. Agent Classification in the LLM Era

4.1 Classification by Interaction Mode

graph LR
    A[By Interaction Mode] --> B[Single-Turn<br/>Conversational]
    A --> C[Multi-Turn<br/>Interactive]
    A --> D[Autonomous]

    B --> B1[ChatGPT single Q&A]
    B --> B2[Code completion]

    C --> C1[Claude Code interactive development]
    C --> C2[Customer service dialogue systems]

    D --> D1[AutoGPT autonomous task execution]
    D --> D2[Devin autonomous programming]
    D --> D3[OpenAI Operator]

4.2 Classification by Capability Dimensions

Dimension	Level 0	Level 1	Level 2	Level 3
Reasoning	No reasoning	Single-step reasoning	Multi-step CoT	Tree search/Reflection
Memory	No memory	Context window	External storage	Hierarchical memory system
Tools	No tools	Single tool call	Multi-tool orchestration	Tool discovery/creation
Planning	No planning	Single-step planning	Multi-step planning	Dynamic re-planning
Collaboration	Single agent	Human-agent	Multi-agent with fixed roles	Dynamic multi-agent

4.3 Classification by Autonomy Level

Inspired by the L0-L5 levels of autonomous driving, agent autonomy can be graded as follows:

Level	Description	Example
L0	No autonomy: pure tool, fully human-controlled	Calculators, search engines
L1	Assistive: suggests actions, human confirms	GitHub Copilot code suggestions
L2	Semi-autonomous: auto-executes subtasks, human supervises	Claude Code edit mode
L3	Conditionally autonomous: mostly autonomous, asks human for help on exceptions	Cursor Agent mode
L4	Highly autonomous: completes full task workflows, only outcome review	Devin autonomous programming
L5	Fully autonomous: no human intervention needed	Theoretical AGI Agent

5. Multi-Dimensional Comprehensive Classification Framework

Agent classification should not be limited to a single dimension but should consider multiple dimensions comprehensively:

\[ \text{Agent Profile} = (\text{Architecture}, \text{Reasoning}, \text{Memory}, \text{Tools}, \text{Autonomy}, \text{Interaction}) \]

For example, Claude Code can be characterized as:

Architecture: Hybrid (LLM deliberation + reactive tool calling)
Reasoning: Level 2-3 (multi-step CoT + self-correction)
Memory: Level 2 (context window + file system)
Tools: Level 2 (multi-tool orchestration: bash, file read/write, search)
Autonomy: Level 2-3 (semi-autonomous to conditionally autonomous)
Interaction: Multi-turn interactive

Summary

Agent taxonomy has evolved from single-dimensional to multi-dimensional, from static classification to dynamic assessment. Understanding different classification frameworks helps with:

Design decisions: Selecting the appropriate agent architecture based on requirements
Capability assessment: Systematically evaluating an agent's various capabilities
Research positioning: Clarifying where research contributions fit in the agent design space
Technical roadmaps: Planning development paths from current to target capability levels

References

Russell, S. & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.
Wooldridge, M. (2009). An Introduction to MultiAgent Systems (2nd ed.). Wiley.
Brooks, R.A. (1991). Intelligence without Representation. Artificial Intelligence, 47(1-3), 139-159.
Muller, J.P. et al. (1995). The Design of Intelligent Agents: A Layered Approach. LNCS 1177.
Wang, L. et al. (2024). A Survey on Large Language Model based Autonomous Agents. Frontiers of Computer Science.

Agent Taxonomy

Overview

Taxonomy Overview

1. Russell & Norvig's Classic Taxonomy

1.1 Simple Reflex Agent

1.2 Model-Based Reflex Agent

1.3 Goal-Based Agent

1.4 Utility-Based Agent

2. Wooldridge's Agency Classification

2.1 Weak Agency

2.2 Strong Agency

3. Architecture Classification: Reactive vs. Deliberative vs. Hybrid

3.1 Reactive Agent

3.2 Deliberative Agent

3.3 Hybrid Agent

4. Agent Classification in the LLM Era

4.1 Classification by Interaction Mode

4.2 Classification by Capability Dimensions

4.3 Classification by Autonomy Level

5. Multi-Dimensional Comprehensive Classification Framework

Summary

References

评论 #