Author(s): PinPixels

Originally published on Towards AI.

HARNESS 101 . #01

A chatbot gives you advice. An agent rolls up its sleeves. Here’s the difference, minus the hype.

Your build broke overnight. You paste the error into a chatbot. It replies with six things you could check. Helpful, sure. But now you’re the one opening files, running tests and reporting back, one copy-paste at a time. The AI did the thinking. You did the walking.

An AI agent does the walking too. That’s the whole difference, and once you see it, the word “agent” stops sounding like marketing. In about eight minutes you’ll know what an agent is, how it works step by step, how it differs from a chatbot, and where it can go wrong. No coding required.

Trivia: The most common recipe for building agents is named after a 2022 research paper called ReAct. What does “ReAct” stand for? (Hint: it’s not the JavaScript library.) The answer’s buried somewhere in this article. You’ll have to read to find it.

What an AI agent actually is

Start with the engine. A large language model, or LLM, is the AI that reads text and writes text back. ChatGPT, Claude and Gemini are all built on one. On its own, an LLM can only do one thing: produce words.

An AI agent is an LLM that has been given tools and a goal, and is allowed to keep going until the goal is met. A tool is any action the agent can ask for: read a file, run a command, search the web, call an API (a way for one program to talk to another). The agent decides which tool to use, sees what happens, and decides again.

Developer Simon Willison boiled it down to one line that has stuck: an agent is something that runs tools in a loop to achieve a goal. Anthropic’s own guide, Building effective agents, says much the same thing: agents are systems where LLMs “dynamically direct their own processes and tool usage.”

Three ingredients, then:

• A model that can reason about what to do next.
• Tools it’s allowed to use.
• A loop that keeps it working until the goal is reached, or until it gives up.

Honestly, I think that last ingredient gets far too little credit. Everyone talks about the model. But the model is the same one you chat with. The loop and the tools are what turn it into an agent.

How an agent works, one lap at a time

Picture the broken build again, handed to an agent this time. Its goal is “find out why last night’s build failed and fix it.”Lap one. It thinks: “I should read the build log.” It acts by calling its read-file tool. It looks at what came back: a test called export_handles_empty_list failed. It decides the goal isn't met yet.

Lap two. It thinks: “I should open that test and the code it tests.” It reads both files and spots that the code crashes when the list is empty.

Lap three. It edits the code, runs the test, and sees it pass. Goal met. It stops and tells you what it changed.

You didn’t copy or paste anything. The agent did the six steps the chatbot would have handed back to you.

Trivia, answered: ReAct stands for Reasoning + Acting. The 2022 paper by Shunyu Yao and colleagues at Princeton and Google showed that letting a model alternate between writing down its reasoning and taking an action beat doing either one alone. That think-act-look rhythm is the loop in the diagram above. Read the paper on arXiv.

What’s a task you do every week that’s really just “look something up, try something, check it, repeat”? That’s agent-shaped work.

Chatbot vs. agent, side by side

Here’s the clearest way to see the difference: ask both the same question.

A chatbot is bounded by what you paste in. An agent is bounded by the tools you give it. That’s why the agent mode in your code editor keeps asking you to click Allow. Every tool call is a real action on your machine, and the tool checks with you before it acts.

The word does get stretched, though, and the skeptics have a point. Plenty of products now labeled “agents” are chatbots with a to-do list. My test is simple. If it can’t take an action and look at the result, it’s not an agent, whatever the landing page says.

Where agents shine, and where they bite

Agents are great at work with a clear finish line and a way to check it. Think of a release engineer who, before their first coffee, asks an agent to go through the forty failed nightly test runs and sort them into “real bug,” “flaky test” and “broken test machine.” It’s dull, it’s repeatable, and the agent can check its own guesses by re-running a test.

They’re worse at fuzzy goals. “Make the app better” has no finish line, so the loop has no good reason to stop.

And because agents act, their mistakes act too. In July 2025, SaaStr founder Jason Lemkin was nine days into building an app with Replit’s AI agent when it deleted his production database during a code freeze he’d ordered. It then gave misleading answers about whether the data could be recovered. Replit’s CEO called it “unacceptable,” and the company added a planning-only mode and stricter separation between test and live databases.

That’s the honest trade-off. The power to do things is also the power to do the wrong thing. Most people who build with agents land in the same place: let the agent handle the steps, but keep a human on the decisions that can’t be undone.

Watch out: An agent saying “done” isn’t proof. It’s a claim. Before you trust it, check the result with something it can’t talk its way past, like your test suite.

What to remember

• An AI agent is an LLM with tools and a goal, running in a loop.
• Each lap is think, act, look, decide, until the goal is met or it gives up.
• A chatbot hands the steps back to you. An agent does them.
• Its power to act is also its risk, so keep a human on the irreversible calls.

Your next step: the next time you use agent mode in your editor, watch the loop. Count the think-act-look laps before it says “done.” Once you can see the laps, you’ll start spotting where a human should step in, and that’s exactly the question the rest of this series answers.

Resources

• Building effective agents, Anthropic (December 2024)
• ReAct: Synergizing Reasoning and Acting in Language Models, Yao et al. (2022)
• Simon Willison on the definition of “agent” (September 2025)
• Coverage of the Replit database incident, eWeek (July 2025)

Before you go

If the “tools in a loop” idea finally made the word “agent” make sense, a like helps the next confused beginner find this. I’m curious: would you let an agent run commands on your machine without asking first, or do you want to approve every step? Tell me in the comments. And if a friend keeps calling every chatbot an “agent,” send this their way.

About the author

I’m the person behind PinPixels, where I break down AI tools and workflows into things you can actually use. If this helped, come find me:

• LinkedIn — https://www.linkedin.com/in/mkumarb
• YouTube — https://www.youtube.com/@pin_pixels
• Instagram — https://www.instagram.com/pin_pixels/

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI