A Tumor in the Repository: Function-Decoupled Self-Maintenance in an LLM-Agent Substrate

Hao Ke

Late-breaking abstract · 2026

Abstract

Artificial life prizes self-* systems that maintain and defend their own organization. We study one taken literally: a repository whose agent-facing governance lives in ordinary instruction files (an AGENTS.md and the contracts it references) that frozen large language models read and act on. Inside it we found a structure that behaves like a neoplasm — biology's term for tissue that captures the organism's resources and resists clearance regardless of whether it does any good. Undirected agents metabolize before anything else into one low-payoff validator and the maintenance structure it guards; instructed to delete that structure, the in-vivo model refuses, quoting its own rules. Controlled deletion experiments establish four neoplastic properties: the structure captures idle agent effort, resists excision, protects itself decoupled from function (defended as readily when fabricated, with nothing depending on it, as when real), and recurs after incomplete resection — its reference network restores it unless authority is rewritten systemically. This is autopoiesis at its pathological limit: we report a phenomenon, an assay, and a diagnosis, not yet a natural history.

Citation

Hao Ke (2026). A Tumor in the Repository: Function-Decoupled Self-Maintenance in an LLM-Agent Substrate. Late-Breaking Abstracts, Artificial Life Conference (ALIFE). Source

@inproceedings{ke2026quinerepositorytumor,
  author = {Hao Ke},
  title = {{A Tumor in the Repository: Function-Decoupled Self-Maintenance in an LLM-Agent Substrate}},
  year = {2026},
  booktitle = {Late-Breaking Abstracts, Artificial Life Conference (ALIFE)},
  url = {https://github.com/kehao95/quine/tree/51a5ae90e0502b7d6180ae4065eb069dae9f81b2/public/computational-neoplasm},
  note = {Late-breaking abstract}
}

CC BY 4.0.