{
  "id": 19177,
  "url": "https://arxiv.org/abs/2608.13228v1",
  "title": "Capability Sheaves for Compositional Agent-Harness Repair: Controlled Quotients and a Real-Repository Stress Test",
  "summary": "Agent harnesses combine retrieval, routing, state, provenance, and verification, but locally successful components may disagree on shared state. We model this failure with a finite \\emph{capability sheaf}: stalks encode typed behavior signatures, restriction maps retain shared fields, and accepted runs are useful global sections. An exact finite constraint-satisfaction problem (CSP) defines acceptance, while a linearized relative cohomology class provides a diagnostic and search feature. A contr",
  "authors": "Saveliy Batruin",
  "category": "research",
  "topics": "healthcare,agents-autonomy",
  "orgs": null,
  "regions": null,
  "published_at": "2026-08-13T13:31:09.000Z",
  "fetched_at": "2026-08-14T05:10:49.168Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/19177",
  "original_url": "https://arxiv.org/abs/2608.13228v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}