Levels of AI Agents: from Rules to Large Language Models

Yu Huang

TL;DR

This paper proposes a structured 6-level hierarchy (L0–L5) for AI agents, inspired by the SAE levels of autonomous driving. It charts the evolution of agentic intelligence from simple rule-based tools to sophisticated, collaborative Large Language Model (LLM)-driven entities capable of autonomous learning and social interaction.

What Problem Does This Paper Address?

The term "AI agent" is currently a catch-all for everything from a simple chatbot to a complex autonomous robot. This lack of a standardized taxonomy makes it difficult for researchers and engineers to benchmark progress or define the "intelligence" of a system.

The author, Yu Huang, addresses this by creating a clear evolutionary path. The paper seeks to answer: How do we categorize the "strongness" and utility of an AI agent? By moving beyond the binary of "it's an agent" or "it's not," the paper provides a framework to understand the transition from symbolic AI to the current era of LLMs and toward Artificial General Intelligence (AGI).

Methods

The paper is primarily a conceptual framework and literature synthesis. It draws on several key AI paradigms to build its 6-level hierarchy:

  1. Symbolic AI & RL: Discusses the limitations of rule-based systems (fragility) and Reinforcement Learning (sample inefficiency).
  2. The LLM Pivot: Identifies LLMs as the new "brain" or controller for agents, enabling emergent abilities like Chain-of-Thought (CoT) reasoning and In-Context Learning (ICL).
  3. World Scope (WS) Integration: Maps agent capabilities against the five levels of "World Scope" (Corpus, Internet, Perception, Embodiment, and Social).
  4. The 6-Level Taxonomy:
    • L0 (No AI): Basic tools with perception and action (e.g., a simple sensor-actuator loop).
  5. L1 (Rule-based): Fixed logic and "if-then" structures.
  6. L2 (RL-based): Learning through trial and error, adding reasoning and decision-making.
  7. L3 (LLM-based): Replacing RL with LLMs, introducing memory and reflection.
  8. L4 (Autonomous Learning): Adding generalization and the ability to learn new tasks without specific fine-tuning.
  9. L5 (Social/Personality): The peak level, featuring emotional intelligence, distinct character traits, and multi-agent collaboration.

Datasets

As a framework and survey paper, it does not introduce new datasets or report on specific training runs. Instead, it references foundational work involving models like GPT-3/4, PaLM, and LLaMA, and benchmarks like the Turing Test and the "Levels of AGI" matrix.

Key Insights and Ideas

  • The SAE Analogy: The most striking idea is using the autonomous driving levels as a template. Just as "Level 5" driving means no human intervention, "Level 5" AI agents represent entities that can navigate social and collaborative environments autonomously.
  • Memory as a Tier-Jump: The transition from L2 to L3 is defined by Memory and Reflection. The author argues that the ability to store experiences and "think about thinking" is what separates traditional RL agents from modern LLM agents.
  • Personality is the Final Frontier: Interestingly, the paper places "Personality (Emotion + Character)" at the highest level (L5). This suggests that true AGI isn't just about raw logic; it’s about the ability to integrate into human social structures through empathy and collaboration.
  • Tool Usage as Generalization: L4 agents are characterized by their ability to generalize tool usage—for example, a model trained to use Yahoo Search successfully transferring that skill to Google Search without retraining.

Key Results

The primary result is the Hierarchical Levels of AI Agents (Table 3), which provides a roadmap for the industry:

  • L0-L2 represent the "Classical" era of AI.
  • L3 is the current state-of-the-art for many "agentic" startups (LLM + RAG + Memory).
  • L4-L5 represent the immediate research frontier, focusing on continual learning and multi-agent systems (MAS).

The paper also highlights the "World Scope" progress: while pure LLMs are stuck at Level 2 (Internet), LLM-assisted agents are currently pushing into Level 3 (Perception) and Level 4 (Embodiment).

Limitations and Future Work

The author acknowledges several hurdles that currently prevent agents from reaching L4 and L5:

  • Hallucination & Reliability: LLM-based agents are prone to making things up, which is dangerous in high-risk domains.
  • Transparency: The "black box" nature of LLMs makes their decision-making processes difficult to audit compared to L1 rule-based systems.
  • Adversarial Vulnerability: LLMs are susceptible to prompt injection and other attacks.
  • Physical Gap: Moving from the "Internet world" to the "Physical world" (Embodied AI) remains a massive challenge in terms of data and safety.

Future work is directed toward improving emotional reasoning, collective intelligence in multi-agent societies, and mitigating catastrophic forgetting during continual learning.

Summary updated

Abstract

AI agents are defined as artificial entities to perceive the environment, make decisions and take actions. Inspired by the 6 levels of autonomous driving by Society of Automotive Engineers, the AI agents are also categorized based on utilities and strongness, as the following levels: L0, no AI, with tools taking into account perception plus actions; L1, using rule-based AI; L2, making rule-based AI replaced by IL/RL-based AI, with additional reasoning & decision making; L3, applying LLM-based AI instead of IL/RL-based AI, additionally setting up memory & reflection; L4, based on L3, facilitating autonomous learning & generalization; L5, based on L4, appending personality of emotion and character and collaborative behavior with multi-agents.
BibTeX
@misc{huang2024levels,
  author = {Yu Huang},
  title = {{Levels of AI Agents: from Rules to Large Language Models}},
  year = {2024},
  eprint = {2405.06643},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.06643},
}
Figure