5-MIN EXPLAINER

What is an agent?

Agent vs chatbot vs copilot vs assistant. The words are abused. Here's the real difference.

What is an agent?

Five years from now, the word “agent” will mean what it actually means. Right now, it’s a marketing buzzword smeared across anything AI-shaped. Let’s unpack the real difference.

The four labels people use interchangeably

You’ll hear “agent”, “chatbot”, “copilot”, and “assistant” used as if they’re all the same thing. They’re not — and the differences matter when you’re choosing how to build.

LabelWhat it actually does
ChatbotResponds. Single-turn or short-turn conversation. No tools, no memory beyond the chat.
AssistantResponds + helps. May have basic tool use (“translate this”, “summarise this”).
CopilotPairs with you in a specific app. Aware of context. Can act inside that app’s surface.
AgentPlans + acts on a goal. Loops: observe → plan → act → observe → repeat. Multi-step. Tool-heavy.

The defining trait of an agent: the loop

A chatbot answers your question. An agent takes your goal and works toward it through multiple steps, deciding each step what to do next.

That decision-making loop is the difference. Without a loop, you have a smart Q&A. With a loop, you have something that behaves agentically.

USER: "Find me three flights to Sydney under $300, then add them to a comparison doc"

CHATBOT  : "Here are some flight search websites you could try."  (responds, done)

ASSISTANT: "I searched and found these flights: ..."             (one tool call, done)

AGENT    : Search → filter → search again → fetch prices → 
           open doc → write rows → verify → report.              (loops)

Why the distinction matters

  • Failure modes are different. Chatbots hallucinate. Agents loop, lose context, take wrong actions, burn budget. Different bugs need different mitigations.
  • Permissions matter. A chatbot reading docs is harmless. An agent with write access is a security surface.
  • Cost is different. Chatbots = one round-trip. Agents = many round-trips (sometimes hundreds).
  • UX is different. Agents need progress visibility, intermediate state, intervention points.

”Agentic” — what it actually means

If you hear someone say “agentic AI”, they could mean any of:

  1. An LLM that can call tools. Loose definition. Honestly, this is just “tool use”.
  2. An LLM that loops. Tighter — it’s making multi-step decisions.
  3. A multi-agent system. Several LLMs collaborating.
  4. Anything AI-related. Marketing. Ignore.

When in doubt, ask: “Does it loop and decide what to do next, or does it just respond?”

The agent stack (mental model)

Every agent has these layers, even if you build them by hand:

┌─────────────────────────────────┐
│ UI / Trigger / Goal-setting     │  ← User interface, API, scheduled trigger
├─────────────────────────────────┤
│ Orchestration / Planner         │  ← What to do next? Often the same LLM.
├─────────────────────────────────┤
│ LLM (the brain)                 │  ← Claude, GPT, Gemini, etc.
├─────────────────────────────────┤
│ Tools / MCP servers             │  ← Read GitHub, write to disk, call APIs
├─────────────────────────────────┤
│ Memory (short / long term)      │  ← Conversation history, persistent state
└─────────────────────────────────┘

That’s it. There’s nothing magical underneath. An agent is just an LLM in a loop with tools and (optionally) memory.

When you don’t need an agent

If your task is single-step, an agent is overkill. Use the smallest pattern that works:

  • One question → one answer? Chatbot.
  • Question + one tool call? Assistant.
  • In-app help? Copilot.
  • Multi-step goal that requires deciding-as-you-go? Now you need an agent.

The biggest mistake people make in 2026 is shipping an agent for what should have been a one-shot LLM call. More moving parts = more things to break.