MARL-SI: Mirror Agent Recursive Learning with Convergence-Guaranteed Self-Improvement

Abstract

Self-improving agent architectures represent a promising paradigm for autonomous capability development. However, deploying these systems in regulated or high-stakes environments requires guarantees of safety, traceability, and stability. This paper introduces MARL-SI (Mirror Agent Recursive Learning), a proposed architectural design focused on safety and trust infrastructure for self-improving agents. The design combines three key properties: (1) a symmetric observer-executor loop where the critic's capability co-evolves with execution capability, addressing the lagging critic problem; (2) a parameter-level validation gate designed to protect against performance regressions by testing proposed changes across tiered quality dimensions; and (3) a version-controlled patch taxonomy to enable rollback and auditability of agent modifications. We also outline a knowledge transfer protocol that provides a fixed meta-observer with operational data while keeping its judgment independent. We present the theoretical framework, including a convergence guarantee based on the Monotone Convergence Theorem, and analyze the design's properties relative to contemporary self-improving architectures. This work is a theoretical proposal, with empirical implementation and validation identified as the critical next step.