What is an AI Agent?

From Single Answers to Autonomous Action

Every AI interaction you've had so far has followed the same pattern: you type something, the model responds, done. That exchange — one input, one output — is the basic unit of prompting.

An AI agent breaks that pattern entirely.

Instead of answering once and stopping, an agent keeps going. It sets a goal, figures out what information or actions it needs, takes steps to get them, observes what happened, and decides what to do next. It loops until the task is finished — or it decides it can't finish.

This is what people mean when they say AI is becoming "agentic." The model isn't just a calculator you query. It's a system that acts.

A Simple Analogy

Imagine asking a junior researcher to find out the current market cap of three companies and summarize the trends.

A chatbot gives you its best answer based on training data — probably outdated, definitely unchecked.

An agent would:

1. Search the web for each company's current stock price
2. Calculate market caps
3. Look for recent news about each company
4. Synthesize everything into a summary
5. Return the result

The agent used tools, made multiple decisions, and produced a result a chatbot simply couldn't.

The Four Core Properties of an Agent

Every AI agent, regardless of framework or complexity, has these four characteristics:

1. Perception

The agent takes in inputs — not just your initial message, but tool results, memory contents, conversation history, and any other context it can access. Perception is about what the agent can see.

2. Reasoning

The LLM at the core of the agent decides what to do next. Given everything it perceives, it plans an action: "I should search for X" or "I have enough information, I'll write the answer now." This is the cognitive step — the model acting as the brain.

3. Action

The agent executes its decision. Actions can include:

• Calling a function or API (search, calculator, database query)
• Writing or running code
• Reading or writing files
• Sending messages
• Calling another agent

4. The Loop

After acting, the agent receives the result and goes back to perception. It sees the new information, reasons again, and decides whether to act further or stop. This observe → reason → act cycle is what makes agents autonomous.

The Agent Loop

Here's the core loop visualized:

Goal / User Instruction
        ↓
   [ Reasoning ]  ← "What should I do next?"
        ↓
   [ Action ]     ← Call a tool, write output, ask a question
        ↓
   [ Observation ] ← See the result of the action
        ↓
   [ Reasoning ]  ← "Did that work? What's next?"
        ↓
  (repeat until done)
        ↓
   [ Final Answer ]

Each cycle is a turn. Simple tasks might take 2–3 turns. Complex research or coding tasks might take 20+.

What Makes a Good Agent Task?

Not every task benefits from an agent. An agent is overkill for "summarize this paragraph." But agents shine when:

A Real Example: Research Agent

Let's say you want a brief report on recent developments in quantum computing.

Without an agent: You'd write a detailed prompt asking for a summary. The model generates text based on its training data, which has a cutoff date. The output might be outdated or miss recent breakthroughs.

With an agent:

Turn 1:
  Thought: I need recent news about quantum computing. I'll search the web.
  Action: search("quantum computing breakthroughs 2026")
  Observation: [search results with titles and snippets]

Turn 2:
  Thought: I see several relevant articles. Let me read the top 3.
  Action: fetch_page("https://...")
  Observation: [full article content]

Turn 3:
  Thought: I have enough information. Let me synthesize a report.
  Action: write_report(...)
  Final Output: [well-sourced, up-to-date report]

The agent used tools to overcome its knowledge limitations and produced a result that's genuinely current.

The Spectrum of Autonomy

Agents exist on a spectrum from barely autonomous to fully autonomous:

Prompt pipeline → fixed sequence of prompts, no branching Workflow → branching logic, but paths are pre-defined Supervised agent → loops with human approval at key steps Autonomous agent → loops without human intervention Multi-agent system → multiple agents coordinate autonomously

You'll often want to start at the supervised end — where you can see what's happening — and move toward autonomy only as you gain confidence in the system's reliability.

Why This Matters for Prompt Engineers

Understanding agent architecture makes you better at prompting in two ways:

1. System prompts for agents need to define not just persona and tone, but tool-use policies, stopping conditions, and how to handle uncertainty.

2. Task decomposition — breaking a big goal into agent-sized steps — is a prompt engineering skill. The better you can structure tasks for an agent, the more reliable it becomes.

Key Takeaways

• An AI agent is an LLM wrapped in a perceive → reason → act → observe loop
• Agents can use tools — giving them access to real-time data and the ability to take actions
• Not everything needs an agent; they're best for multi-step, open-ended tasks
• The core skill is understanding how the loop works — the frameworks change, but the architecture doesn't
• Next lesson: the four internal components that make up any agent