LongHarness paper reports 4.4% to 61.1% exact match gain on OfficeQA Pro V2 via custom harness

A paper introduces LongHarness, a model-agnostic agent harness that keeps evidence and open conclusions as explicit state instead of one growing chat history.

The authors report that GLM-5.3-Flash rises from 4.4% to 61.1% exact match on OfficeQA Pro V2 using the same model. The harness is built for corpora of 100M tokens and more, and the authors report that LongHarness beats Claude Code by 12.1 points on V1 and Codex by 18.9 points on V2. Elvis Saravia (@omarsar0) describes the paper on X, and the figures are the paper authors' reported results.