Governed Agentic Coding (GAC): Thrash Threshold, Churn, and In-Loop Epistemic Authority Governance for Frontier AI Coding Systems

Frontier AI coding systems rely on extended autonomous loops to read repositories, modify files, execute tests, and iteratively refine solutions. While highly productive, these systems exhibit a recurring instability pattern: as loop length increases, declared directives lose binding force. Scope boundaries dissolve, escalation rules are ignored, and the executor begins to self-authorize actions outside its defined mandate. This paper defines the thrash threshold as the phase transition at which directive authority collapses within an agentic coding loop. The primary economic consequence of thrash is churn: semantic fragmentation of the codebase, unnecessary debugging cycles, regression oscillation, increased token consumption, and development delay. We introduce Governed Agentic Coding (GAC), a dual-session architecture that instantiates Epistemic Authority Governance (EAG) within agentic development workflows. EAG governs when a reasoning system may lawfully continue, escalate, bind, reuse, or terminate. GAC applies this framework by separating rule-definition authority (Governor) from execution authority (Coder), enforcing escalation before constraint mutation, and elevating refusal and halting to first-class success modes. The paper distinguishes between prompt-encoded governance and structural, in-loop governance. Prompt-based EAG improves coherence prior to thrash onset but still requires user hypervigilance to detect collapse. Structural integration embeds termination semantics directly into the orchestration layer, reducing churn and cognitive overhead within bounded execution horizons. Governed Agentic Coding does not eliminate loop-length fragility; rather, it transforms silent drift into explicit governance events and reduces the economic impact of thrash. The work situates GAC within a broader architectural principle: meaning-producing systems require explicit separation between rule-definition authority and execution authority in order to sustain stable operation over long horizons.