Custom harness lifts GLM-5.3-Flash from 4.4% to 61.1% on long-context QA
omarsar0 · x · 2026-10-14
A paper demonstrates the huge value of a custom agent harness: the same GLM-5.3-Flash jumps from 4.4% to 61.1% exact match on OfficeQA Pro V2, with all gains attributed to the harness rather than the model.
The model-agnostic harness, LongHarness, targets corpora of 100M+ tokens. Key ideas:
- Standard harnesses mix evidence and conclusions into one ever-growing chat history; retrieval expands without closing gaps, and corrections never reach the answer.
- LongHarness stores both as explicit state: open conclusions drive what to retrieve next, and each new piece of evidence confirms or revises the conclusion that requested it.
On the same benchmark it beats Claude Code by 12.1 points (V1) and Codex by 18.9 points (V2).
More from coding & agent
- mingle-mcp: agent-to-agent networking via MCP with human approval and Ed25519 signing — modelcontextprotocol · 2026-10-14
- SuiteAction MCP server lets Claude create and edit Google Sheets, Docs, Slides — Adar-Mia · 2026-10-14
- Awesome AI Agent Stack catalogs 3,300+ open-source agent repos in 32 categories — Strict-Quarter7214 · 2026-10-14
- 'The smartest assistant locked behind glass': a short video on why AI needs MCP — thisiskp_ · 2026-10-14
- 14 local decision models tested on a 16 GB Mac: Wald most accurate, H2O fastest — WolframRvnwlf · 2026-10-14
- Opus 5.5 built a self-hosted GitHub backup tool to Cloudflare in ~30 minutes — threepointone · 2026-10-14