Corpus‑level long‑context reasoning requires agents to acquire and integrate evidence from document collections that exceed their context windows. Existing agent harnesses allow incremental access to such corpora, yet the acquired evidence and evolving conclusions remain tangled in an ever‑growing interaction history. This implicit state management leads to two coupled failure modes: unbounded exploration—retrieval expands the history without closing decision‑relevant information gaps; and evidence‑state drift—previously gathered evidence or later corrections fail to influence the current conclusion. LongHarness introduces a model‑agnostic framework that makes the interaction between evidence and conclusions explicit. Its core mechanism, Evidence‑Conclusion Coevolution, uses unresolved conclusions to steer retrieval toward missing information, while newly acquired evidence strengthens or revises the conclusions that guided the search. In parallel, Evidence Lifecycle Management keeps accumulated evidence current and traceable throughout the interaction. Together, these mechanisms focus retrieval on remaining information needs and keep conclusions aligned with up‑to‑date evidence. Experiments on OfficeQA Pro V1/V2, AA‑LCR, and LongBench v2 employ GLM‑5.3‑Flash and Qwen3.8‑Flash. Compared with direct reasoning using the same backbone, LongHarness raises GLM‑5.3‑Flash exact match on OfficeQA Pro V1 from 15.8% to 60.2% and on V2 from 4.4% to 61.1%. Against state‑of‑the‑art agent harnesses, it outperforms Claude Code by 12.0 points on V1 and Codex by 18.9 points on V2. With Qwen3.8‑Flash, LongHarness reaches 61.7% on V1, surpassing Claude Code by 8.3 points and Codex by 12.0 points.
Review