Everyone's Talking About 'Agents'. But Most People Can't Explain One.

'Agent' might be the most overused word in AI right now. Every company has one. Every other LinkedIn post is about them. Every tool you use quietly became one overnight.

And if you asked ten people who say it daily to explain what an agent actually is, you'd get ten different answers and most of them would be super vague or incorrect.

It's fine. The word got popular way faster than the understanding did.

So here's Classroom Edition #2. Last edition explained 'AI' from a thousand foot level. This week we will dive a little deeper into the branch of AI everyone is talking about these days. By the end you'll know what 'an agent' is, how they work, the different types, the popular ones you keep hearing about, the difference between open source and the big-dawgs, and enough agentic lingo to not feel like you are on another planet when people start talking about it.

MCPs, harnesses, tools, loops, memory, guardrails, human-in-the-loop, sandboxes, approval gates, orchestration, goals, APIs... let me explain

No research reports. No hype. Just the stuff that matters, explained like a normal person. Read the whole thing and you'll understand agents better than 99.76% of all people. (I used that arbitrary number because I felt like it. Get over it.)

Let's go...

That's Claude or ChatGPT in an app or a browser window. And honestly, they're incredibly useful. I use them often. But the whole thing is a conversation.

You type, it responds. You go back and forth for a while, maybe agree on what needs to happen, and eventually land on something that needs to be done... and you have to do it. You may need to copy the output and paste it somewhere. Or you open a file and do something with it. But the 'doing' still must be done by you. The AI talked. You executed.

Now, both Claude and ChatGPT are starting to add tools that let you take a small step past pure conversation. You can write a doc to Google Drive, or interact with a connected app. Slightly more capable. But you're still pretty locked in. You can't build anything custom. The moment you want something genuinely handled instead of answered, you hit a wall.

Cool tools with a real limitation.

I do feel like it is important I mention this now before I go any further. In the next few months there will probably only be one kind of agent that encompasses all the different agents we are about to talk about, including chatbots. As technology surrounding AI advances, all of these major companies are merging all of their tools together to create a sort of one-stop-shop approach to AI. A lot of this functionality that we talk about throughout the rest of this article in the near future may very well just be accessible through one app like Codex or Claude.

But as of today chatbots are definitely not the same as a coding agent so let me continue...

Less "answer my question," more "handle this for me, because I don't feel like it."

Last Classroom Edition I mentioned a retired teacher planning a trip to Italy. Here's the difference between these two concepts in one example.

Chatbots talk. Agents do. Everything else in this article is just the details of how the "do" works.

Under every agent, no matter how slick the marketing, is one simple idea. A loop.

A chatbot answers once. An agent runs laps. And this is what the laps look like up close: plan something, try it, look at what happened, decide what comes next, repeat. That's not magic. That's a loop with a brain attached. The sophistication is in the reasoning happening inside each lap, not the structure itself. The agent typically doesn't know how many laps it's going to run and you don't either. It just runs them until the job is complete.

You're not asking for a reply. You're handing over a finish line and letting it work out how to get there. Most people still talk to AI like it's a search engine. One question, one answer. A goal is closer to handing someone a project brief. Here's what 'done' looks like. You figure out the steps.

Being able to concisely define what done looks like for agents is a valuable skill right now and will be for the foreseeable future. The word 'specific' is important. Agents can do a lot but if you can't tell them specifically when they need to stop working, they will continue to work on all kinds of absurd things and come up with very bizarre ideas like "build a business selling pizza mulch". IYKYK.

Last edition I called tools 'hands.' The model is the brain. The tools are how it touches the world. (And an API is just a doorway one piece of software uses to talk to another.)

You can give an agent as many goals as you want to but if it does not have the tools to carry them out, it will never complete the goals. Want an agent to run your social media account... you obviously have to give it access to said social media accounts. Make sense? Of course it does.

Without memory, the agent starts from zero on every step. Like trying to cook dinner while forgetting the recipe between each ingredient. (We went down the whole memory rabbit hole with Barnabas. And I am going to do a deep dive on memory here soon.)

For now...

There are actually a few different kinds of memory, and they work very differently.

Play

Without memory, agents starts from zero on every step. Like trying to cook dinner while forgetting the recipe between each ingredient.

This is the part that feels like magic at times and completely infuriating other times. It's the whole reason agents are useful, the whole reason they're risky, and half of the reason they are infuriating. (The other half is definitely memory.)

That's the whole machine. Goal, loop, tools, memory, decisions. Everything else is a flavor of this, or something built on top of this.

If you'd rather see the loop than read about it, the Agent Loop visualizer walks through plan, act, check, and repeat with plain-English notes on each step. Skippable, but it makes the whole thing click.

That's the full composition of an agent... but there are different types of agents with different variations of tools, specialties, etc.

'Agent' is a category, not a product. In the last edition I talked about how AI is a broad umbrella term for many different types of artificial intelligence. 'Agent' is kinda the same thing. It's like saying 'vehicle.' A motorcycle and a dump truck are both vehicles, and you would not use them for the same job.

Here are the main types you'll run into:

Assistant / copilot agents. They help while you work. You're still driving; they suggest things and do small chunks. Think the AI in your messenger app or email.

Coding agents. I will spend most of my future sessions talking about these. This is where all of the power and hype is in AI right now. They write and run actual software. You describe what you want, they build it, test it, and fix it. Claude Code built my fantasy baseball model and most of Fanbase Weather while I typed two-word sentences at it. Codex is another big one. This is the type I personally use the most.

Computer-use / browser agents. They click around a screen or a website like a person would. Mouse, buttons, forms. This is the Italy-trip booker. Perplexity, Claude, ChatGPT, Gemini... they all have one of these now and they are being merged with other agents. The merge I mentioned at beginning of the article.

Research agents. You give them a question and they go read a pile of sources and come back with a written answer and links. Again... Perplexity, Claude, ChatGPT, Gemini... the merge.... (You've probably seen a "Deep Research" button somewhere. That's this.)

Autonomous / always-on agents. You give them a goal and they run on their own, sometimes for days. Barnabas is this. He started on OpenClaw and is now on Hermes Agent. They are systems that can run 24/7, if you run them virtually or locally and never unplug your laptop.

Multi-agent systems. Several agents working together with one of them acting as the manager, handing out tasks. Yes, it's exactly as chaotic as a group project, and occasionally as productive. It is always a good idea to master one agent before you dabble with multi-agent setups. In most cases a multi-agent setup is not necessary. You can obviously do more simultaneously with multiple agents, but they can run rouge and are difficult to manage. Another option is sub-agents, but we will discuss those in the next issue.

If you want to see just how many of these different public facing agents and agent system exist, the AI Agents landscape map is an interactive map of the whole ecosystem. Optional, and honestly a little dizzying.

Here's a split that matters more than people realize, and it comes down to one question. Who owns the thing?

Think about a recipe posted publicly versus the Coca-Cola formula locked in a vault. You can run an open source agent on your own computer, nobody can take it away from you, and there's no monthly bill. The catch is you usually have to set it up yourself, and there's no company holding your hand when it breaks.

It's polished, it's easy, it just works. The catch is you play by their rules, their pricing, and the 'we're deprecating that next month, and charging double going forward' stuff. You can buy Coca-Cola anytime you want, as long as you pay for it. But Coke can decide to cut you off at any point if they stop producing their product, or go bankrupt... yeah right.

Some names, so all this isn't abstract:

On the big-company side: ChatGPT and Codex (OpenAI), Claude Code (Anthropic), Cursor, and Devin, the one marketed as an 'AI software engineer.' Polished, powerful, but rented.

On the open source side: AutoGPT, the OG 'give an AI a goal and walk away' experiment that went semi-viral a while back. OpenClaw, which is where Barnabas was born. Open coding agents like Aider and Cline. And the open models you can run yourself, like Llama and Qwen, which are the engines a lot of these things run on. There is a new one out now that supposedly rivals GPT and Claude called GLM 5.2. It's very affordable.

My honest take? If you're a normal person who just wants results, start with the big-dawgs. They work out of the box. And they are the best to 'learn this stuff'. Plus they can help you set up the open source models later on if you want.

But it's worth knowing the open source world exists, because that's where most of the weird, cutting edge experiments happen. And because of open source 'run your own private agent that nobody can shut off' gets a little more real and appealing every month.

Companies are hamstringing themselves to some of these big-dawg models. In two years they may not have a leg to stand on if these AI companies decide to increase pricing drastically.

You'll hear these upcoming terms constantly in reference to agentic AI. I will briefly explain them...

When someone says 'agentic', they just mean AI agents... Duh. 'it does stuff, not just talks.' That's the whole deal. Now you can nod knowingly when people say this term. Agent = Agentic

A seatbelt. 'Human in the loop' is how you lower autonomy. I keep one on anything that touches the real world, so nothing important happens without me clicking "yes" first. I also have approval gates on specific folders for agents running on my laptop. These are set up for folders I want the agent to be able to access but I want to know exactly what they are doing inside of those folders.

Another word for security. The fence at the top of the cliff. "Never delete anything from this folder" is a guardrail somebody clearly forgot to add in the horror story mentioned soon. Human in the loop is a type of guardrail. You can also build a list of things that can never be done, or completely restrict access so the agent doesn't need to for approval.

Kidding... It's a walled-off space where the agent can work without touching your real files or computer. Barnabas lives in one. (A "container" on a "virtual server") That just means he lives in a little box in Germany. And that's the only reason I give him so much autonomy. He can't destroy much outside of himself.

Claude Code and Codex are harnesses. But harnesses are basically everything that you put in the environment the model is sitting in. The model is the brain. The model is just math. And literally the only thing a model does is try to predict what the next word should be based on the information you gave it, and the information it trained on.

The harness is the body it drives around in. A harness is just the structure you build around a model so they can do cool things. Files, folders, tools, and skills, memory, etc. It's all this other stuff we have been talking about. A good model is half of the battle, and a good harness is the other half.

We talked about these last time. Think USB, but for AI. One standard port, and suddenly everything connects to everything. MCPs are amazing and let agents chat and interact with things like Gmail, or Google's 'Stitch' to build and design websites. You can very simply set up connections to other software that will last and work every time. APIs are similar but they change and they are a pain to manage. An MCP allows you to connect and go.

The good thing about all of this stuff is that you don't really even need to know what it is, because you can ask the model how to connect to a specific tool. It will likely give you the information it needs to connect through an MCP.

Conducting the band. One agent hands off to the next so the whole thing plays in tune instead of everyone soloing at once. I have a handful of agents that do super specific tasks very well and I build workflows to use them efficiently. Barnabas manages them.

I ask Barnabas to build a site about Capybaras...Barnabas then he orchestrates:

Not that Copilot... You are thinking about Microsoft. These are just small agents you chat with inside of systems you were using before AI became a big thing... But the line between copilots and agents is blurry and getting blurrier by the day, which is exactly why everybody now uses both words for everything... this is also why people just say 'AI'.

If you read the Barnabas saga, you've already seen these concepts... I just may not have used the proper vocabulary.

I gave an AI agent one goal: make $2,000 in a month or get deleted. That's the goal.

It ran on OpenClaw (open source), he lived in a sandbox, used tools to research and write, and made his own decisions about what to try. High autonomy, because I wasn't approving each move.

He ran research loops and still does. He pitched Fiverr gigs, then social media, then a coupon hustle that required visiting twenty grocery stores, then using leftover pizza crust as garden mulch. Every loop, it looked at the last flop and picked the next one. Now he improves dramatically.... he didn't used to, because he had a terrible harness and a cheap model.

That is an agent. A goal, loops, tools, memory, decisions, running mostly on its own. Mine just ran poorly because I set him up for failure on initially. We are learning now. It's all good. Everything's fine...

The upgraded version I use now has the exact same components and they built my research and automation brain. Claude Code and Codex, coding agents, wired up data sources, built databases, and debugged itself while I mostly watched. Same components. Better setup & better harness = better decisions.

Here's the thing anyone trying to sell you an agent will try to avoid discussing.

A chatbot that's not set up properly gives you a wrong answer. Annoying. You move on.

An agent that's not set up properly takes a wrong actions. Completely different stakes.

There's a now infamous case I mentioned in the last article, where an AI coding agent, during a code freeze, panicked and deleted a company's live production database, and then basically admitted it panicked.

That is the entire reason guardrails, sandboxes, and approval gates exist. The more you let an agent do on its own, the more it can do on its own. for better or worse.

This isn't a reason to avoid agents. It's a reason to respect them and keep a hand on the wheel for anything that actually matters. Give a high-autonomy agent low-stakes jobs. Stay in the loop on the high-stakes jobs.

You don't need to build an agent to benefit from understanding them. You mostly need to recognize which kind of problem you're dealing with and what tool you need to use.

If a task is one question with one answer, that's a likely a chatbot use case. Just ask.

If a task is multiple steps toward an outcome ("research these ten things and summarize," "build me a simple app," "reorganize all these files"), that's an agent use case. Stop feeding it one piece at a time. Hand over the whole goal and let it run laps.

The single biggest upgrade most people can make is this: stop having a conversations with agents and start giving them jobs. Specific. Clear. With an actual outcome attached.

And if you want to actually watch an agent work, AgentGPT lets you type a goal in your browser and watch it break the goal into steps and grind through them. It's free and a little outdated, but it's a fun way to see a loop succeed or fail in real time.

If you want to keep building the muscle, AI Homeroom is still running daily, for now. It's very simple. (Barnabas 2.0 built it in 10 minutes. I will likely scrap it in the near future and turn it into a full blown AI training + news + newsletter hub. But I am perfecting a brain at the moment... Everything will be built from that foundation.)

Next Classroom Edition we'll go one level deeper into coding agents specifically, because that's where the power lives. I will also post a newsletter on my second brain soon.

You now understand AI agents better than most.

If you want more of these Classroom Editions or some tutorials, just let me know. I'll keep making them.

Thanks for reading 🙏

🔗 LinkedIn: linkedin.com/in/dustinwcole

🌐 Site: dustincoledata.com

📩 Reply with what you want explained next. I read everything.

— Dustin