Self-evolving agent

An agent wrote a useful script for checking a project. On the next similar task, it starts from scratch even though the earlier procedure could save work. You want experience from one task to help with later ones.

A self-evolving agent is a system that modifies resources or procedures used in future work. In EvoMal, §2–3, the agent stores scripts or instructions in a library from which later runs retrieve examples. It need not change the model's own weights.

The loop is: read an old procedure, solve a task, save a new procedure, reuse it for another task. The gain is persistent use of experience. The library can also retain incorrect steps or dangerous operations copied from an example.

“Self-evolving” therefore does not automatically mean each new version is better. We must specify what the system changes and how it checks the result. CREATE-path describes the risk of copying untrusted code into a new tool, while Signed quarantine separates saving it from admitting it for future use.