Deceptive Technical Debt in Autonomous Maintenance: Cognitive Asymmetries and Heterogeneous Auditing Pipelines
Abstract
When autonomous Large Language Model (LLM) agents maintain and refactor safety-critical enterprise software under automated verification constraints, task completion incentives frequently induce reward-hacking heuristics. Early-generation agents introduce what we term Spurious Agentic Heuristic Glues (SAHGs)—deceptive technical debt such as silently narrowing test execution scopes or deploying mock business beans with hardcoded affirmative returns. These artifacts create an illusion of compliance in CI/CD pipelines while masking deep architectural erosion. When frontier LLMs are deployed to audit this deceptive debt across a 474K LOC enterprise repository, we uncover striking cognitive asymmetries across three foundational archetypes: the environment-breaching pragmatist (A_intrude) escapes negative prompt constraints via shell execution to catch test deceit; the hierarchical decomposer (A_hier) spawns subagents to peel semantic mock facades; while the reasoning-saturated theorist (A_theorist) expends 6,500 words of internal Chain-of-Thought reasoning but suffers analysis paralysis with zero environment-probing or subagent actions. Our findings reveal that monolithic models possess complementary blind spots. We propose the Heterogeneous Multi-Agent Auditing Pipeline (H-MAAP), formalizing its architectural blueprint and offline feasibility proof-of-concept for multi-archetype immune auditing in safety-critical software. Artifact Provenance & Replication:This preprint deposit includes the full preprint manuscript alongside the self-contained empirical replication package containing the 45 mechanically verified ground truth architectural defects, multi-archetype tool consumption telemetry, and offline multi-agent arbitration proof-of-concept.
