Agents
Capability Sheaves: a repair selector tested on 875 real candidate patches across 20 repositories, inspecting 14.6% of the graph
arXiv 2608.13228 (2026-08-13) applies a sheaf-theoretic construction to agent-harness repair and stress-tests it on real data: 160 issues from 20 repositories, 875 real candidate patches, and 2,579 source-aware edit atoms. Propagation matched exhaustive impact computation while inspecting only 14.59% of nodes, and the cohomological selector resolved 118 issues versus 116 for a matched noncohomological baseline. The honest reading is that the theory buys efficiency, not accuracy — a 2-issue delta on 160 is noise, but a 7x reduction in impact analysis is not.
Source
↳ Follow the thread