0:01 Thousands of companies claim to sell AI agents. Gartner estimates only about 130 of them actually do. A lot of the rest are chatbots, assistants and automation tools they already sold, with a new label. There's a name for that. Agent washing. So when your boss asks for an agent, or a vendor pitches you one, you need to know which kind you're getting. Get it wrong, and you either pay for an agent and get a script, or build an agent where a script would have done.
0:36 One question tells them apart, and you can ask it of anything. Who decides what happens next?
0:51 A model with tools
Here's the plan. A model with tools. The loop that makes it an agent. Then how to tell an agent from a workflow, and how good agents really are. Start with the model on its own. Text in, text out. It can't open a file, or call your API. So you give it tools. A short list of functions it's allowed to ask for, each with a name and a description of what it does.
1:20 Now, instead of an answer, it can write a request. Read sales dot CSV. That request is called a tool call. Your code reads the file, and hands back what's inside. That's one step. And it isn't an agent yet.
1:41 What makes it one is a loop
What makes it one is a loop. You give it a goal, and run that step again and again. Every time, the model decides what to do next. Say you ask it why sales dropped last month. Lap one. It asks what's in your data folder. Your code runs that, and back comes a list of files. Lap two. It picks the sales file and asks to read it. Back come the numbers. Lap three.
2:11 It spots one region that fell off a cliff, so it asks for that region's orders. Lap four. It has enough. So instead of a request, it writes the answer: sales fell because EMEA orders dropped. That's how the loop ends. The model stops asking for tools. Or, if it runs too long, your code stops it. Nobody wrote that sequence. The model chose each step after seeing what came back from the one before.
2:44 So, who decides what happens next? In an agent, the model does. Four pieces make that work. A model that chooses. Tools that let it act. A context that holds everything so far. And a loop around the outside. Everything around the model is called the harness. Tools and context each get an episode of their own. Anthropic, the company behind Claude, puts it in one line.
3:14 Agents are typically just LLMs using tools based on environmental feedback in a loop.
3:24 Why bother with a loop?
So why bother with a loop? Here's the same job without one. A workflow: a script with three fixed steps. Read sales dot CSV. Compare each region with last month. Report the biggest drop. It's cheaper, it's faster, and it does the same thing every time. Then someone renames the file to Q four sales. The workflow goes looking for sales dot CSV, doesn't find it, and stops.
3:57 The agent reads the folder, sees a file that looks like sales figures, and carries on. That's the whole difference. If you wrote the order of the steps, it's a workflow. If the model picks the order, it's an agent. Here's how to answer that question for any system. Cross the model out. If the order of the steps is still written down in what's left, it's a workflow, however much AI is inside.
4:26 If all that's left is a loop with no one to choose, it's an agent. A simple chatbot fails the test instantly. One message in, one reply out, nothing chosen. Even Open AI's guide to building agents says it. Simple chatbots are not agents. But let that chatbot call a tool, look at what comes back, and choose what to do next, again and again, and it passes.
4:54 That's how the big chat apps work now, when they search the web or run code. The chat window doesn't decide it. The loop does. The same goes for workflows. Handing a model tools changes nothing if your code still picks the next step. But if one step gives the model a goal and lets it loop until it's done, that step is an agent. Plenty of real systems are a workflow with an agent inside.
5:24 And a plain workflow is fine. Most software should be one. The trouble starts when you call one an agent. When it breaks, you tweak the prompt, but the fault is in the steps your code wrote.
5:40 How good are agents, really?
So how good are agents, really? A group called METR measures exactly that. They time how long tasks take a skilled person, mostly software tasks, then find the length of task an agent can manage on its own. That length has doubled about every seven months since twenty nineteen, and faster lately. At the end of 2024, that was tasks of about half an hour. By 2026, the best models were past a full working day.
6:14 The catch is what manage means. It's succeeding about half the time. If you need four times out of five, the tasks have to be about five times shorter. So the line is climbing fast. But a good score on a test still isn't a job you'd hand over and walk away from.
6:38 So that's an agent
So that's an agent. A model in a loop with tools, choosing its own next step, working towards a goal. The loop itself is about forty lines of Python. That's the easy part. The hard part is everything a demo never shows. Tool descriptions the model actually understands. A context that doesn't fill up with junk. And step fourteen, when a tool hands back garbage, and the agent carries on anyway.
7:09 Because one bad step can quietly poison every step after it. That gap, between a demo and something you'd trust with real work, is what this channel is about. No hype, and no agent washing. Just how these things actually work, where they break, and how to build them so they don't. And you can check every source. We start with the model itself. That's next.
