Case study demonstrates continuous autonomous workflow execution across hot-swapped language model engines, indicating that file-based harness architectures preserve operational identity.
The Harness Craft: Identity on Disk, Engines as Parts. An agent is a character on disk, not a process in memory. Identity, doctrine, working state and protocol live as plain files — markdown state-machines, capability cards, hook guards — and the language model driving the agent is a hot-swappable part, injected with that identity at every session birth. This is argued not as design philosophy but measured as operational fact: a publication desk's engine was swapped mid-life between vendor model families and kept publishing peer-gated records without losing a turn; an accidental engine drift was caught by the transcript, not a dashboard; and the brakes that held were the ones bound to process, not to any particular engine. Part of a series; it cites the companion campaign paper (10.5281/zenodo.22086013) and the related process-and-memory reports in the same corpus. Bilingual: full text in English and Turkish. AI-transparency note: the draft was written by an AI agent under human direction (the human authors held go/no-go authority), blind cross-family jury-reviewed (producing model family excluded from the bench), and reviewed against the source ledgers by the human authors before publication.
No takes yet. Share an insight, caveat, or question.
Melikoğlu et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: