The question of how language models should manage their own context—what to remember, what to discard, when to update—has traditionally been treated as an external problem, something a harness or prompt engineer solves for a frozen model. This work inverts that assumption, giving models native control over context as an editable file and asking what they learn to prioritize. The result surfaces a productive tension: models understand context better than they generate it, yet here they're asked to *manage* it dynamically, which blurs that boundary and raises questions about whether intrinsic context curation changes what comprehension even means. The multi-agent implication is equally open—if contexts can coexist as files and be edited by the models themselves, does that sidestep the coordination constraints that plague larger agent collectives, or does it simply defer them to a new layer? And when CLMs are steered via natural-language instructions and reinforcement learning, are we training models to be better context managers, or training ourselves to articulate what good context management looks like—and might those two things diverge?
Lines of inquiry this post opens
Not questions with answers — ways of approaching this material across the collection.
How do standardized protocols improve multi-agent coordination and reliability?