Every consultant has a way of doing things. How you open a discovery call, which red flags you look for in a contract, the shape of a weekly client report, the words you never use. For most of a career that lived in your head and in a folder of old documents. Agents change that, because an agent is where that know-how gets written down in a form that does the work.
That makes a good agent valuable in a way a prompt never was. A prompt is a sentence. An agent you’ve used for six months is a record of every time it got something wrong and you told it how you’d actually do it. Nobody else has that record. It’s yours, and it compounds: the agent is better in month six than in month one because of the corrections you put into it.
Where lock-in happens
Here’s the problem. The easiest place to build an agent is inside a chat product. You make a Custom GPT in ChatGPT, a Project in Claude or a Gem in Gemini, you paste in instructions, upload some files, and it works. Then over the following months you tweak the instructions in that settings box, add files, maybe connect a few tools.
Six months later, ask yourself a few questions:
- Where is the history of what you changed and why? Usually nowhere. The settings box holds only the latest version.
- Could you run the same agent on a different model tomorrow? Only by copying and pasting, and anything built on that product’s features doesn’t come with you.
- If a better model ships from a different company, how long until you can use it? For most people the honest answer is “I’d have to rebuild.”
That’s lock-in, and it’s rarely the model’s fault. Models are getting easier to swap every month. What locks you in is where you kept your work.
The idea of this course
Keep the agent in files you own, and treat each provider as a place to run it.
The files are plain text: a set of instructions, a few skills, the knowledge it needs and the tools it can call. They live in a folder on your machine, ideally in git, so every change has a date and a reason. Claude, ChatGPT and Gemini all read these files directly or through a short setup step, and by the end of this course you’ll have done that setup for all three.
Two things follow once you work this way. The first is that you can use whichever model is best for a job. In my own setup, research agents run on the strongest model I can get, and coordination agents run on a faster, cheaper one. When a new model comes out I can try it on one agent in an afternoon. The second is that improvement becomes a habit instead of a project. A correction is a one-line change to a file, and the next run picks it up everywhere that agent runs.
Do this
Pick one agent you already rely on, or one task you’d like an agent to take over. Write down in a sentence or two what it does, where it lives today, and what you’d lose if that product disappeared tomorrow. You’ll rebuild it as a portable agent over the rest of the course.