Answers are grounded in this paper and cite their sources.
Explores structural degradation in mature software systems due to automated code changes, suggesting critical implications for quality management.
Large language models are increasingly deployed as autonomous code-generation agents operating on mature production codebases. This paper identifies a structural degradation process — the agentic decoherency tumbler — in which repeated machine-generated modifications cumulatively erode the internal coherence of a software system while leaving standard quality metrics undisturbed. The mechanism operates as follows. An agentic session begins from a default state that does not include the system's full constraint web — the accumulated interdependent decisions that define why the code is shaped the way it is. Without iterative alignment between the model's statistical defaults and the system's specific structural requirements, each generated modification drifts toward the median of the model's training distribution. That drift is small on any single pass. But each modified codebase becomes the input for subsequent agentic sessions, shifting the local statistical center and increasing the probability of further drift on the next pass. The process is self-accelerating: the tumbler's output feeds back as its input. The resulting degradation has several components: convergence of structurally diverse code toward homogeneous median patterns, loss of local optimizations that reflected problem-domain asymmetries, expansion of code volume without corresponding expansion of capability, accumulation of implementation decisions with no provenance or rationale, and displacement of the human constraint-web knowledge required to detect or reverse the process. This degradation is invisible to the metrics that drive organizational decisions — velocity, test pass rates, headcount efficiency — and becomes visible only through lagging indicators that appear after the structural damage is done. The process is irreversible by the same mechanism that produced it, because the agent is the tumbler. No claims are made about LLM cognition or intent. The problem is entirely mechanical, arising from the interaction between how token prediction works, how context windows constrain what the model can observe, and how iterative application of a median-seeking process to a feedback loop produces convergent degradation.
No takes yet. Share an insight, caveat, or question.
Geoffrey Howland (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: