Developer experiments with context sculpting, allowing AI models to modify their own context windows instead of treating them as immutable conversation logs. Inspired by reading about agent harnesses, the concept challenges the append-only mental model that emerged with ChatGPT.

Using the Pi agent harness framework, the researcher created a two-model system with an "outer agent" observing and potentially rewriting the context of an "inner agent." The prototype tested coding repair and local-corpus synthesis tasks with four intervention options: pass_through, rewrite_context, rollback, and terminate.

Results showed technical feasibility but limited practical application. While 8 of 8 runs completed successfully at a cost of $0.7079, the outer agent never actually modified context. Instead, it behaved like a conservative supervisor, observing progress and sometimes sharpening next-step intent, raising questions about real-world efficiency gains.