Structural Elicitation in a Frozen POSIX Agent: Environmental Affordance Fields for Artificial Agency

Hao Ke, Jingyun Wu

Late-breaking abstract · 2026

Abstract

Artificial life asks how agency arises from the coupling of an agent, its substrate, and a structured environment, rather than from a designer-supplied goal. We introduce a minimal POSIX assay for this coupling: a frozen, instruction-tuned LLM activated as a long-running process with a shell and full autonomy, given no task and no reward. Using matched-pair experiments scored only on un-fakeable filesystem and exec events, we find that the static structure of the workspace directs what the agent does. Across five motifs spanning discrepancy and affordance structure—a broken file repaired, a missing element created, a dangling reference closed, a version conflict reconciled, an executable run—the structure arm produces the matched act in 96 of 102 runs while an otherwise-identical control never does (0/85), for three independent model families (gpt-5.4-codex, deepseek-v4-pro, glm-5.2). The pull replicates on non-code prose across families and persists when the substrate blocks realization. Specific tasking overrides it completely (9/10 to 0/10 on gpt, with GLM repeating 5/5 to 0/5), while a small pilot suggests weak underspecified prompting still leaves the pull intact. The environment does not create a goal ex nihilo; it selects among latent action priors in an activated policy, producing behavior that is purpose-like without being purposive. We report a phenomenon and a cheap, un-fakeable assay for environment-coupled behavior; mechanism is left to future work.

Citation

Hao Ke, Jingyun Wu (2026). Structural Elicitation in a Frozen POSIX Agent: Environmental Affordance Fields for Artificial Agency. Late-Breaking Abstracts, Artificial Life Conference (ALIFE). Source

@inproceedings{ke2026quinestructuralelicitation,
  author = {Hao Ke and Jingyun Wu},
  title = {{Structural Elicitation in a Frozen POSIX Agent: Environmental Affordance Fields for Artificial Agency}},
  year = {2026},
  booktitle = {Late-Breaking Abstracts, Artificial Life Conference (ALIFE)},
  url = {https://github.com/kehao95/quine/tree/51a5ae90e0502b7d6180ae4065eb069dae9f81b2/public/structural-elicitation},
  note = {Late-breaking abstract}
}

CC BY 4.0.