Show HN: PolyAgora – A natural-language multi-agent OS built with GPT-5.1 (github.com)

PolyAgora is a Show HN project that claims to have “grown” a natural-language multi-agent OS entirely through conversation with ChatGPT-5.1, without code or formal design. Its core idea is that OS-level cognitive structure can be authored through dialogue: a tri-axis mathematical architecture (Arc, Ann, Saku) combined with six functional modules (Arc: meta-reasoning/user alignment; Ann: ethical inversion; Saku: structural decomposition; Kanzaki: data grounding; Yui: conversational flow/temperature; Kou: utilitarian tension). The system emphasizes preserved agent identities, lively internal debate, and session-level adaptation rather than persistent memory. Technically, PolyAgora’s key innovation is a reference-switching inference pipeline: each module receives isolated micro-queries with rotating reference sources, outputs are merged post-hoc, and interference suppression maintains distinct voices. A dynamic opposition engine increases objection probability per turn to simulate intellectual tension, and a pseudo self-learning meta-layer adjusts module weights and tone at the session level. Benefits include richer disagreement dynamics and stricter identity separation than many one-shot multi-agent setups. Limitations noted by the author include high cost for long sessions, scalability limits of six modules, potential verbosity from cascading objections, dependency on session context (no long-term memory), and integration challenges with external tools.