What is an AI Agent?

Gentle introduction to Agentic Systems

An AI Agent refers to a system or program that is capable of autonomously performing tasks on behalf of a user or another system. An Agent combines a Large Language Model (LLM) with the ability to take actions. The LLM provides reasoning, while tools and APIs enable the actions.

When you give an agent a task, it:

1. Uses the LLM to understand what needs to be done.
2. Selects appropriate tools from its available set.
3. Takes actions using those tools.
4. Evaluates the results.
5. Repeats this cycle until completing your task.

There are different degrees to which systems can be agentic. A system is more "agentic" the more an LLM decides how the system can behave.

Agentic vs. RAG

AI architectures like Agentic and RAG (Retrieval-Augmented Generation) leverage LLMs differently to complete tasks.

Agentic Architecture: This approach uses an LLM to autonomously decide and act based on the task's needs, utilising tools and APIs for execution. It's dynamic, adjusting actions based on results, which promotes continuous improvement. Characterised by its ability to initiate and learn from interactions, the Agentic architecture supports proactive and adaptive decision-making.

RAG Architecture: Conversely, RAG enhances LLM responses by integrating information retrieved from a large data corpus. This method is suited for tasks requiring detailed knowledge, as it augments the model’s output with precise, retrieved information, ensuring depth and accuracy.

Each architecture uses LLMs to suit different needs: Agentic for autonomous operations and RAG for depth of information, highlighting their distinct applications and dynamics. Also Agentic architectures will incorporate RAG setup to dynamically add context to it's prompts.

There are several reasoning paradigms for solving multi-step problems. Let's explore a couple.

Reasoning Paradigms

ReAct (Reasoning and Action)

In this paradigm, you can instruct agents to "think" and plan after each action is taken and with each tool response to decide which tool to use next. The ReAct framework defines how AI agents think and act. Each step follows this sequence:

1. Think: The agent reasons about what to do next.
2. Act: The agent uses a tool or takes an action.
3. Observe: The agent processes the results.

This cycle repeats until the agent completes your task. ReAct makes agent decisions visible by showing each thought step, similar to Chain-of-Thought prompting.

How to Implement ReAct:

1. Structure your prompts to request explicit reasoning steps.
2. Configure the agent to display each thought before taking action.
3. Ensure the agent updates its context after each observation.
4. Allow the cycle to continue until reaching the desired outcome.

ReWOO (Reasoning Without Observation)

In this paradigm, you instruct agents to plan all actions upfront before execution, rather than reacting to each tool output. The ReWOO framework defines how AI agents plan and execute tasks efficiently. Each module follows this sequence:

1. Planning: The agent analyzes your prompt and creates a complete action plan.
2. Execution: The system runs all planned tool calls and collects outputs.
3. Response: The agent combines the plan with tool outputs to generate results.

This approach reduces computational overhead and lets you verify the plan before execution. ReWOO makes agent decisions efficient by planning everything at the start, avoiding the need to reason after each tool output.

How to Implement ReWOO:

1. Structure your prompts to encourage comprehensive upfront planning.
2. Review and approve the agent's action plan before execution.
3. Execute all tool calls in a single phase.
4. Generate the final response by combining the initial plan with tool outputs.

Types of Agents

Simple Reflex Agent

An agent that converts current inputs directly into actions using predefined rules, without storing any history.

Model-Based Reflex Agent

An agent that maintains an internal model of its environment to make decisions based on both current and past observations.

Goal-Based Agent

An agent that plans action sequences to achieve specific objectives by evaluating different paths to its goals.

Utility-Based Agent

An agent that selects actions by calculating which outcomes provide the highest value across multiple criteria.

Learning Agent

An agent that improves its performance over time by updating its knowledge base with new experiences and feedback.

Here is a comparison of the different types (Substack doesn’t allow tables, hence the screenshot from Obsidian).

It is a foregone conclusion that AI agents will reshape how we approach complex tasks, blending reasoning and autonomous actions to achieve the desired outcome. Whether through step-by-step reasoning like ReAct or upfront planning with ReWOO, these frameworks can unlock new possibilities . By tailoring agent types—reflexive, goal-driven, or learning—to specific needs, you can better leverage their capabilities to solve problems and achieve meaningful outcomes.