Playing NetHack with a Coding Agent: A Logged Case Study of Long-Horizon Agency under Human Guidance

This paper presents a retrospective case study of a human-guided coding-agent workflow that completed NetHack 5.0.0. The agent developed a terminal scaffold, consulted documentation and source code, maintained persistent state records, and continued across repeated interruptions. Native game records confirm ascension after 23,008 game turns, with 1,177,800 points and 146/176 hit points. The campaign archive contains 242 controller sessions, 46,230 input requests, and 32 reconciled recovery records.

The operator explicitly permitted save restoration and narrowly scoped selection of favorable random outcomes, including modifications to previously generated state. The study therefore analyzes an assisted completion rather than an unassisted benchmark result. It reconstructs the architecture, gameplay trajectory, intervention policy, and selected failures, distinguishing final-state progress from discarded work and outcome verification from performance replication.

The supplementary archive contains LaTeX sources, analysis scripts, computed metrics, selected native completion evidence, recovery and intervention records, and integrity manifests. It does not contain the complete gameplay archive. AI assistance is disclosed in the manuscript.