After years of work behind the scenes, my wife recently launched KelleyKouture, a high-end fashion female footwear company.  Like any new business, it needs a hundred things done, across a dozen different systems, in a particular order, every single day.  I have been building with AI long enough to know what it can and cannot do.  So I created her something that lets it handle the coordination, without her needing to open a chat window and type instructions every five minutes.

What I built, is something the internet is currently calling an agent framework. ooh fancy!

Here is the thing: once you strip the label off, it is almost embarrassingly simple.

Scroll through GitHub right now and you will find projects with hundreds of stars, breathless README files, and architecture diagrams that look like they were drawn by someone who charges $400 an hour.  Words like “agentic pipelines,” “multi-modal orchestration,” and “skill-based execution layers” are everywhere.  Consultants are building decks around this.  Vendors are pricing it like infrastructure.

Take three high quality projects currently doing the rounds on GitHub.  

• Open Montage
  Produces full videos, via scripts, calling a variety of APIs and utilizing ffmpeg
  
• Open Design
  An open source alternative to Claude Design and Google Stitch for building a UX
  
• career-ops
  Suite of skills to automate your job search, preparing tailored resumes/cover letters, including applying for the job for you online

Three completely different domains.  Strip them back and the pattern is identical every time. And those are just three.  There is OpenClaw, Hermes, a dozen others with names that sound like they were chosen specifically to make you feel like you need a briefing before you can touch them.  Different names, same pattern every time.

• Skills are markdown files.  Plain text.  Instructions written in English that tell the AI what to do in a given situation.  Not code.  Not configuration files with obscure syntax.  A text file you could write in Notepad.
  
• Tools are scripts.  A Node application.  A Python function.  A Linux utility like ffmpeg already sitting on your machine.  Things that do one specific job when called.
  
• Memory is a file.  A JSON object, a markdown document, a YAML config.  Something that holds state between sessions so the agent remembers what happened last time.

That is the whole framework.  Skills, tools, memory.  The AI reads the instructions, calls the right script, writes back to the file.  Repeat.

An agent is just a fancy word for a macro.  Or a task.  Or a workflow.  Whatever word your industry used before this one came along and started appearing in conference talks.

The buzzwords are doing a specific job here, and it is not the job you think.

They are pushing you back.  They are making you feel like this requires specialists, budget sign-off, and a three-month implementation project.  They are making you feel like you need to understand the architecture before you can touch the architecture.

You do not

If you can write down what you want to happen, in order, you can build one of these.  Anyone who has ever documented a process already knows how to do the hard part.  The rest is telling the AI which text file to read next.

The barrier to entry is not a computer science degree.  It is a text editor.

Before you bring in a consultant to implement an agentic workflow for your business, open a blank document and write down exactly what you want the system to do.  Step one, step two, step three.  What information it needs.  What it should do with the result.  Where it should put the output.

You just wrote a skill file.  You are already most of the way there.

The people selling complexity around this are, almost without exception, the people who benefit from your confusion.  The technology is not complicated.  The problem you are trying to solve is the complicated part.  That was always true.  AI just gave it a new name.

The barrier to entry for agent programming is not a computer science degree. It is a text editor. The “agent developer” is anyone who can write down a process and knows which folder to put the file in. Does that sound complicated?

Let me prove it

Here is a complete, working agent skill you can run right now in Claude Code. Create a folder called .claude/skills/commute/ in any directory on your machine and drop in a single file called SKILL.md with this content:

# Commute Check
You are a commute assistant.  When this skill is invoked, start by asking
the user one question:
"Are you driving or taking the train today?"
Wait for their answer, then check current travel conditions between:
FROM: 1111 E Broad Street, 3rd Floor, Richmond, Virginia 23219
TO: 11800 W Broad St, Richmond, VA 23233
Use web search to find current traffic or service conditions for whichever
mode they chose.  Return a plain English summary.  No more than four lines.
Lead with the estimated travel time.  Flag anything worth knowing, like an
accident, roadworks, or a service delay.

Now open Claude Code in that directory and type: /commute

Claude reads the skill, asks you whether you are driving or taking the train, waits for your answer, then goes and checks.  One file.  One command.  The AI holds a conversation and acts on what you tell it.

That is a macro.  That is also, apparently, an “agent framework”. Basic one yes, but still an agent framework.

Congratulations.  You are now an agent developer.  Go and update your resume and LinkedIn profile.

Now, one thing worth being honest about: what you just built is reactive.  It does something when you tell it to do something.  You invoke it, it responds.  That is still enormously useful, but it is not the whole picture.

Frameworks like Hermes and OpenClaw add a little sugar on top.  They invoke skills on a schedule and push the result to you, an email, a WhatsApp message, a Signal alert.  Your commute check could run automatically at 4pm every day and land in your pocket before you have even thought about leaving the office.  That sounds more impressive than it is.  I will show you how those work in a follow-up post, and I promise you, it is painfully stupidly easy. Hint you don’t even need to install one of these to get a lot done.

For now, go create a folder and write a text file.  The rest, as they say, is just details.

AI Disclaimer: Gemini Nano Banana Pro was used to generate the photo – from the 1994 Dumb and Dumber movie.