Governed Framework for Controlled Proposal-Based Modifications in Autonomous Agent Systems

Recent advances in autonomous agent systems enable iterative self-improvement through proposal-based modifications to code, prompts, and execution strategies. However, existing systems lack structured mechanisms to regulate these modifications, resulting in instability, unsafe updates, and limited traceability. This work introduces a governed supervisory framework that mediates all modification proposals through a deterministic lifecycle comprising evaluation, decision-making, controlled execution, verification, and rollback. The architecture enforces separation between proposal generation and decision authority, ensuring that all changes are subject to explicit governance constraints. The system is implemented as a modular, research-grade framework with a sandboxed execution environment and append-only logging for traceability. Evaluation under a controlled experimental setup demonstrates increased selectivity in proposal acceptance, reduction in rollback occurrences, improved decision consistency, and enhanced stability of lifecycle outcomes. These observations indicate that structured governance mechanisms enable more reliable and controlled handling of proposal-based modifications without altering proposal generation behaviour.