Your agent runs the same job a thousand times and ends every one exactly as capable as it started.
The improvement loop runs through you, by hand, in the prompt.
Every mistake the agent makes becomes another instruction in the file.
It grows, and you find out it got too long when rules start being skipped without warning.
The ticket got reopened. The deploy held. The refund went out twice.
That signal sits in your logs and never reaches the thing that needed to know it.
Everyone wants a smaller model tuned on their own traffic.
You need the input, the action taken and what happened next, joined up. Most stacks keep the output and throw away the two things that give it meaning.
