What is an AI agent? Architecture and concepts

Section 10.7.1 : What is an AI agent? Architecture and concepts

🎯 Learning objective

Understand what an AI agent is, how it works (perception-reasoning-action loop), and why agents represent the next revolution after chatbots. You'll learn to distinguish different types of agents, understand their internal architecture, and identify use cases where an agent is more suited than a simple chatbot.

From chatbot to agent: a conceptual leap

A chatbot (ChatGPT in classic conversation mode, Claude) answers your questions one by one. It is reactive — it waits for your input, produces a response, and stops.

An AI agent is fundamentally different: it is proactive. It receives a goal (not just a question), plans the necessary steps, uses tools (web search, calculations, code execution, email sending...), observes results, adapts if something doesn't work, and continues until the desired result is achieved — all with minimal human intervention.

Analogy: a chatbot is like an employee who answers the phone and only responds to questions asked. An agent is like an autonomous employee to whom you say "organize next week's conference" — and they plan, book the room, send invitations, prepare materials, and tell you when it's done.

The 4 components of an AI agent

1. The LLM (brain): the language model that reasons, plans, and decides which actions to take. GPT-5, Claude Opus 4.6, Gemini 3.1 Pro are the most commonly used brains for agents in 2026. The model choice determines reasoning quality and planning capability.

2. Tools (hands): the agent's action capabilities. Each tool is a "function" the agent can call:

• →Web search (to find current information)
• →Code execution (to calculate, analyze data, create charts)
• →File reading/writing (to manipulate documents)
• →API calls (to interact with external services: email, CRM, calendar...)
• →Web navigation (to visit sites, fill out forms)

3. Memory (context): what the agent "remembers" during its mission:

• →Short-term memory: conversation history and recent action results
• →Long-term memory: persistent knowledge (RAG, databases)
• →Working memory: the current plan, completed steps, remaining steps

4. The reasoning loop (autonomy): the mechanism that allows the agent to:

• →Evaluate its progress toward the goal
• →Decide on the next action
• →Execute the action
• →Evaluate the result
• →Adjust the plan if necessary

AI agent architecture

The ReAct loop (Reasoning + Acting)

Most modern agents follow the ReAct pattern, which explicitly alternates between reflection and action:

Concrete example — the agent receives: "What is our main competitor's revenue in Q4 2025, and how does it compare to ours?"

1. →Thought (reflection): "I first need to identify the main competitor. I'll search our internal database."
2. →Action: Call to CRM tool → query "main competitor"
3. →Observation: "The CRM indicates that the main competitor is TechCorp Inc."
4. →Thought: "Now I need to find TechCorp's Q4 2025 revenue. This info is probably public if it's a listed company."
5. →Action: Web search → "TechCorp Inc revenue Q4 2025"
6. →Observation: "According to the annual report, TechCorp made $45M in Q4 2025."
7. →Thought: "Now I need to find our own Q4 2025 revenue for comparison."
8. →Action: Call to internal database tool → Q4 2025 sales query
9. →Observation: "Our Q4 2025 revenue = $38M."
10. →Thought: "I have both figures. I can calculate the gap and produce the report."
11. →Final Answer: "TechCorp achieved $45M in Q4 2025, 18.4% more than our revenue of $38M. The gap narrowed compared to Q3 (where it was 25%). Sources: TechCorp annual report, internal CRM."

Why ReAct is powerful: the model "thinks out loud" at each step. This allows:

• →Breaking down a complex problem into manageable steps
• →Choosing the right tool for each step (not a single tool)
• →Verifying consistency before continuing (self-correction)
• →Providing traceable reasoning (the agent's logic can be audited)

Types of agents in 2026

Agents in daily life: you're already using them

You're probably using agents without knowing it:

• →ChatGPT with Code Interpreter: when you upload an Excel file and ask for analysis, ChatGPT writes Python code, executes it, observes the result, adjusts if needed → that's an agent
• →Perplexity Deep Research: when you ask a complex question, Perplexity plans multiple search queries, consults dozens of sources, synthesizes → that's a research agent
• →Microsoft Copilot: when you ask "summarize this week's important emails and plan my replies," it accesses Outlook, analyzes, prioritizes, drafts → that's an agent
• →Claude Code: when a developer says "add unit tests for this function," Claude reads the code, understands the logic, writes the tests, runs them, fixes errors → that's a code agent

Risks of autonomous agents

The more autonomous an agent is, the higher the risks:

Risk 1 — Amplified hallucination: a chatbot that hallucinates produces a bad answer. An agent that hallucinates can send an erroneous email, modify a file with false data, or make a decision based on flawed reasoning.

Risk 2 — Irreversible actions: an agent that deletes a file, sends an email, or makes a purchase can't easily "undo."

Risk 3 — Infinite loops: a poorly designed agent can get stuck in a reasoning loop without ever reaching the goal.

Best practices:

• →Human-in-the-loop: the agent asks for confirmation before irreversible actions
• →Budget limits: cap the number of iterations and API cost per execution
• →Complete logging: record every thought/action/observation for audit
• →Sandbox: test agents in an isolated environment before production