Belief-Calibrated Optimization makes the coding-agent-as-optimizer state its hypothesis before editing

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv 2609.01861 (2026-09-01) targets the now-common loop where a coding agent reads scores and traces and iteratively edits the scaffold around a frozen model. Each edit rests on a belief about how the environment will respond, and that belief is normally implicit, living in the agent's reasoning on the current call. The paper makes the world model explicit and calibrates it, which is directly applicable to anyone running a self-improving harness loop against benchmark scores.