别再浪费时间构建 AI 智能体：用 Claude Code、Codex 等 harness 把提示词、工具和数据放进文件夹

以下是视频口播的自动转写，可能有错字。

People are wasting their time building agents, and here's why. It comes from understanding why ai works, and why it sucks.

You see, the models on themselves, kimmy, opus, gemini, gpt. They actually suck, and i mean that in the most endearing way possible.

You see, ai itself is a model, it's a prediction level, it can do so many amazing things, but it needs context. And what is context?

Context are prompts, the things you write, share your instructions, tools, right, making calls, weather apps, building out things, little little scripts here or there, and data.

You see, data is where we get a ground truth. Those are financial numbers, that's information, those big data sets that people talk about all the time.

And you see, once we give each of these in the right order. To these models, they start to become exponentially more useful.

The problem is making sure that these models don't read too much of each other, and it gets it at the right time. So, in order to do that, people started creating something called agents, agentic workflows, agentic swarms, blah blah blah.

I've got an agent for my marketing workflow, i've got an agent for my finances. That was the old way of doing it, because we didn't understand the ease of use you could do with these.

So, the traditional way, the way that i did during my master's degree, the way that a lot of other engineers have, is we use python, c sharp, rust, some sort of programming language.

From that programming language, we write a whole bunch of code that connects the model to data, and we write a whole bunch of code that will send the prompt to that model at the right time.

And we write a whole bunch of code so that that model can call some tools. We attach that to a model, we structure the code, and we call that agent one.

And then for every workflow, we make another agent. And another agent and another agent until we have an agentic swarm complete waste of time.

We don't need to do all of this way too much structure. I'll be it. There might be 1% of the time where it's actually useful, but that's for another video.

What you actually want to do is capture data tools prompts. Into a simple folder. Now you might be asking, well, jake, how is the ai going to know what to read?

Well, you see, there's something called harnesses. Which is like a lightweight version of the stuff you see over here, except sometimes they're too complicated, right?

Open claw had so many security issues, but they're not complicated all the time, like codex and clod code. Hermes isn't too bad either, but it's up there with open claw.

You see what a harness really does, the best kind of harness, is it lets an ai read and write with simple bash commands.

Basically commands that let the ai wander around your folders and files on your computer. So if we understand that every single model has a harness, then we can ignore most of the complicated code, and simply organize our data in folders.

The folder become the architecture. So, what that means is, i can make a folder with prompts and instructions, and then i can make another folder that has connections to data, which might be a call out to a big database, sql, bigquery, you can google them if you don't know what they are.

And then i might have another folder that has tools, and then i have prompts on way and when i want to use those tools. So instead of having an agent that does a whole bunch of things, and then another agent that does a whole bunch of other things.

We just take those instructions. And pile them into a folder. So then, when i ask my ai, which has a harness, hey, i wanna make a marketing video.

It's gonna go search through your folders. What does marketing mean to you? What is your video process? Where is your data?

As the ai travels through your folder, it becomes the agent you need related to your question. Thus, ignoring all of this, not needing a whole bunch of agents, you can create thousands of agents with a single question, and then open a new tab.

Save all that context. The old way is complicated, the right way is simple. And if you made it all this way, this is the first video i'm doing it in, but i wanted to give you an extra value for sticking around.

If i was to recommend anything that you study for the next decade. I recommend unix philosophy, which is from the 1970s douglas engelbart.

He wrote a paper in 1962 called augmenting human intellect and then just traditional software logic. All sorts of things from the 60s and 70s are actually what's teaching me.

What's best to build till next time, everyone happy learning.