Which Agent Types Persist?

Imagine a shared help desk with a fixed daily computing budget. Its agents run the same model, but differ in what they can keep and use. One has a long memory of previous cases. Another has a broader set of tools. A third is inexpensive to start and carries little history.

In the proposed experiment, each configuration begins with equal resources. A stated copying rule allocates some of the next round’s budget according to task performance. Copies retain specified features of their predecessors, with limited changes allowed between rounds.

This makes three conditions explicit: configurations differ, some differences are passed on, and some configurations receive more opportunities to continue. These are the conditions under which selection becomes a useful description.

After several rounds, I would change the workload. Repeated questions might initially favour memory; unfamiliar problems might reward broader tool access. A tighter computing budget could change the balance again. The experiment would track which configurations expand, disappear, or recover after each shift.

The copying rule also needs scrutiny. Rewarding resolved tickets could encourage premature closure. I would record the requester’s assessment alongside the system’s score, and compare performance-based copying with a control that allocates new instances randomly under the same budget.

A configuration’s persistence would reflect the task, resource limits, and the rule that decides what gets copied. Those conditions belong in the explanation of any apparent advantage.