The Modular Monolith as AI-Agent-Native Architecture: Containing Autonomous Code Erosion at 475K LOC under GxP Compliance

Abstract

As Large Language Model (LLM)-driven autonomous coding agents rapidly reshape software engineering, the physical topology of software systems faces fundamental reassessment. While microservices long served as the industry standard for managing human communication bottlenecks, empirical evidence indicates that when individual engineers or small agent clusters act as the primary coding workforce on mission-critical systems, distributed microservices introduce an acute structural dilemma: the "Single-Agent Scope Trap". Network boundaries shatter symbolic semantic continuity, inducing parameter hallucinations, covert mock fallbacks, and non-deterministic transaction tearing. Concurrently, unconstrained monolithic architectures rapidly decay under the shortcut biases of LLMs. To resolve this architectural impasse, we present an extensive empirical investigation of the AMF-LIMS platform—a 474.9K LOC regulatory life-sciences system operating under FDA 21 CFR Part 11 and national pharmacopoeial mandates. Spanning 346 real-world multi-agent development sessions, 37 industrial Git milestones, and 114,357 tool invocations, we formulate and validate the AI-Agent-Native Modular Monolith paradigm, establishing shift-left proactive containment (pre-generative classpath blindness and in-flight type pruning) backed by Spring Modulith and sub-second deterministic bytecode gateways. Our quantitative findings demonstrate that: (1) deterministic compile-time and bytecode hard gates intercepted 1,244 code erosion events, achieving a 92.69% self-healing convergence rate with a +67.48 pp net suppression advantage over soft prompts, suppressing production defect escapes to zero (Defect Escape Rate, DER_prod = 0.0%); (2) dismantling seven cyclic deadlocks optimized the Robert C. Martin Main Sequence Distance by 85.9% (D = 0.054); (3) a paired 20-task natural quasi-experiment with in-situ telemetry confirmed significant causal error reduction (W = 0, p = 8.85e-5 via Wilcoxon Signed-Rank Test), ruling out maturation threats; and (4) the modular monolith reduced resident memory by >80% (1.91 GB), accelerated cross-domain calls to 0.012 ms (>1,500x speedup), compressed agent task context by 94.7% (21.8k Tokens), and elevated first-pass code generation (Pass@1) from 41.2% to 89.6%.