Welcome to your second lesson on building AI agents. In our previous discussion, we established the fundamental distinction between simple LLM calls, deterministic workflows, and true controller-based agents, framing the latter within the formal structure of a Markov Decision Process. This laid the groundwork for a more architectural understanding of agentic systems.
Today, we will build upon that foundation by developing a systematic way to classify agents. You'll learn to analyze and categorize any agent architecture along four principal axes: its degree of autonomy, its access to external tools, its planning horizon, and the number of actors involved. Understanding this multi-dimensional design space is critical for an engineer, as it dictates an agent's capabilities, complexity, and suitability for a given task. These are not rigid boxes but spectra, and an agent's profile across these dimensions defines its character.
The Four Axes of Agent Classification
An agent's architecture can be characterized by its position along four key spectra. Let's explore each one in detail.
1. Autonomy: The Spectrum of Independence
Autonomy refers to the extent to which an agent can make decisions and operate without direct human supervision. This ranges from simple automation to sophisticated self-governance. A highly technical paper by Händler, "A Taxonomy for Autonomous LLM-Powered Multi-Agent Architectures," provides a useful framework for grading autonomy.
[PDF] A Taxonomy for Autonomous LLM-Powered Multi-Agent Architectures
This paper introduces a formal taxonomy for classifying agents. We'll focus on its definitions of autonomy and how it's balanced with alignment.
First, in Section 2.2, familiarize yourself with the high-level system characteristics, particularly the concept of Balancing Autonomy and Alignment. This introduces the core tension in agent design. Next, read Subsection 3.1.1, "Autonomy", to understand the definition. Then, focus on the "Autonomy Levels" description that follows. It provides a clear, three-tiered classification: L0: Static Autonomy: The agent's operations are heavily constrained by predefined rules. L1: Adaptive Autonomy: The agent can adjust its operations within a framework set by the architect. L2: Self-Organizing Autonomy: The agent can dynamically learn and tailor its operations in real-time. Finally, review Subsection 3.1.2, "Alignment," to understand how user-guidance and real-time feedback serve as constraints or steering mechanisms for the agent's autonomy.
A key mechanism that enables higher levels of autonomy is self-reflection. Instead of just executing a task, an agent can be prompted to critique its own output and iterate. Andrew Ng provides a clear example of this principle.
What's next for AI agentic workflows ft. Andrew Ng of AI Fund
In this clip, Andrew Ng discusses "reflection" as a design pattern that allows an agent to improve its own work, demonstrating a step up in autonomy.
Watch the segment from so reflection. Notice how the LLM is prompted first to generate code, and then, in a separate step, to critique that same code, finding bugs it initially created. This self-correction loop is a hallmark of more autonomous systems.
An agent with L2 autonomy might not only self-correct but also independently decide to learn a new skill or modify its core strategy to better achieve a long-term goal, moving beyond simple task execution.
2. Tool Access: Extending Capabilities into the World
An LLM's core capability is text generation. Tool access is what transforms it from a pure text processor into an agent that can act upon the world. This axis describes the variety and sophistication of external tools an agent can use, such as APIs, code interpreters, databases, or even robotic actuators.
The "Environment" block in the CodeAct diagram perfectly illustrates the concept of tool access. It's an interface layer that allows the agent to:
- Seek Information: Use tools like web search to overcome its knowledge cutoff.
- Use Software: Execute code, perform calculations, or run data visualizations.
- Access Memory: Query databases or other external knowledge stores.
- Interact with the Physical World: Control robots or other hardware.
Andrew Ng also emphasizes the importance of tools in expanding what an LLM can do.
What's next for AI agentic workflows ft. Andrew Ng of AI Fund
This segment provides an overview of how tools, from web search to code execution, augment an agent's abilities.
Watch the clip from The second design pattern. Ng points out that tool use is a fundamental technique for gathering information, taking action, and even interacting with modalities the LLM doesn't natively support, like image manipulation.
The expressiveness of an agent is directly tied to the power of its tools. An agent that can only call a weather API is far less capable than one that can write and execute arbitrary Python code, as seen in the CodeAct example. Designing robust, well-documented tools is a central task in agent engineering.
3. Planning Horizon: From Simple Reactions to Complex Strategies
The planning horizon defines the complexity of tasks an agent can tackle. It's the difference between a simple reactive system and one that can formulate and execute a complex, multi-step strategy over time.
- Short-Horizon Agents: These agents excel at single-step tasks or very short sequences, like answering a question by calling a single tool. Their "plan" is often just a single action.
- Long-Horizon Agents: These agents can decompose a high-level, ambiguous goal into a sequence of concrete steps, execute them, and adapt the plan based on intermediate outcomes.
This ability to plan is what elevates an "agentic workflow" above a simple, non-agentic prompt.
What's next for AI agentic workflows ft. Andrew Ng of AI Fund
Here, Andrew Ng contrasts a simple, one-shot prompt with a multi-step agentic workflow and shows an example of an agent creating a complex plan.
First, watch the introduction from today the way, where he contrasts the "type the essay from start to finish" approach with an iterative agentic plan: outline, research, draft, reflect, and revise. Then, jump to the section on planning from planning. The example given (generating an image in a specific pose) is broken down into a multi-step plan involving several different tools. This decomposition is a core function of planning agents.
The paper "AI Agents vs. Agentic AI" formalizes this distinction. It notes that simple "AI Agents" handle single, specific tasks, whereas more advanced "Agentic AI" can manage complex, multi-step tasks requiring coordination. The architectural viewpoint for this in the Händler paper is Goal-Driven Task Management, which involves Decomposition, Orchestration, and Synthesis of results—the key ingredients of planning.
4. Number of Actors: Single Agents vs. Multi-Agent Systems
The final axis considers whether the system is a single, monolithic agent or a society of collaborating agents. While a single agent can be powerful, complex problems often benefit from a "divide and conquer" strategy where specialized agents collaborate.
A recent paper provides an excellent conceptual split between these two paradigms, labeling them "AI Agents" and "Agentic AI."
AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and ...
This paper draws a clear line between single-agent and multi-agent systems, providing a rich vocabulary for their comparison.
First, read the abstract to get a high-level summary of the distinction. An AI Agent is a modular, task-specific system, while Agentic AI involves multi-agent collaboration. Next, read the section "Conceptual Leap". This explains why the field is moving towards multi-agent systems to overcome the limitations of single agents. Finally, review Table I under the section "Key Differentiators". This table succinctly contrasts the two paradigms across features like Task Complexity, Collaboration, and Applications.
This multi-agent approach is not just a theoretical construct. Systems like ChatDev, mentioned by Andrew Ng, simulate a team of software developers.
What's next for AI agentic workflows ft. Andrew Ng of AI Fund
This final clip illustrates a powerful multi-agent collaboration pattern where different agents adopt specific roles to collectively solve a problem.
Watch the segment on multi-agent collaboration from multi-agent collaboration. The ChatDev example, where an LLM is prompted to act as a CEO, designer, product manager, and tester, shows how role-playing can be used to structure complex, collaborative problem-solving.
This "society of mind" approach, where specialized agents collaborate and even debate, often yields more robust and comprehensive solutions than a single agent could produce alone. This is particularly relevant for your interest in AI tutors, which could be designed as a multi-agent system (e.g., a "Socratic Questioner" agent, a "Curriculum Planner" agent, and a "Knowledge Assessor" agent).
Synthesizing the Axes: Agent Profiles
These four axes are not independent. An agent's design represents a specific point in this multi-dimensional space. The Händler paper visualizes this idea using "system profiles" in the form of radar charts, which map an agent's capabilities across various architectural aspects. These aspects are direct correlates to the axes we've discussed:
- Autonomy: The levels (L0-L2) are plotted directly.
- Tool Access: Represented by
Integ(Resource Integration) andUtil(Resource Utilization). - Planning Horizon: Represented by
Decom(Decomposition),Orch(Orchestration), andSynth(Synthesis). - Number of Actors: Represented by
CommP(Communication Protocol),NetM(Network Management), andAGen(Agent Generation).
By examining these profiles, you can quickly grasp the architectural strategy of a given agent system, such as Auto-GPT or MetaGPT, and understand how it balances autonomy, tool use, planning, and collaboration. This analytical approach is invaluable for reverse-engineering existing systems and designing new ones.
Conclusion
In this lesson, we have moved from a general definition of an agent to a structured framework for its classification. You've learned to analyze agents along four critical dimensions: autonomy, from static rule-following to dynamic self-organization; tool access, which gives agents leverage over their environment; planning horizon, which separates simple reactive agents from complex strategists; and the number of actors, distinguishing lone workers from collaborative systems.
These axes provide a powerful vocabulary for discussing, comparing, and designing agent architectures. As you progress, you will see all advanced agentic patterns—from ReAct to multi-agent debates—as different ways of navigating this design space.
In our next lesson, we will zoom back in to the core of a single agent. We'll dissect the canonical perceive → reason → act → observe loop and identify precisely where the essential components—the LLM controller, tools, and memory—fit within this cycle. This will set the stage for understanding the powerful reasoning patterns that drive modern agents.