A new agent harness called LongHarness claims to fix a basic flaw in how AI agents read huge document piles: they forget what they found.
Researchers built LongHarness as a model-agnostic harness for AI agents reasoning over document corpora exceeding 100 million tokens, far more than any model can hold in context at once. Its two components, Evidence-Conclusion Coevolution and Evidence Lifecycle Management, track what evidence has been found and what conclusions it supports, instead of letting that information get buried in a sprawling interaction history. Tested on benchmarks including OfficeQA Pro V1 and V2, AA-LCR, and LongBench v2, the harness took GLM-5.3-Flash's exact-match accuracy on OfficeQA Pro from 15.8% to 60.2% on V1 and from 4.4% to 61.1% on V2, compared with the same backbone reasoning alone. It also beat Claude Code by 12.0 percentage points on V1 and Codex by 18.9 points on V2, and with a different backbone, Qwen3.8-Flash, reached 61.7% on V1 - 8.3 points ahead of Claude Code and 12.0 ahead of Codex.
The gains swing wildly across tests - roughly 4x on one OfficeQA Pro version, nearly 14x on another - which says more about how punishing these benchmarks are for agents with no memory discipline than about any single multiplier. The real finding here isn't the percentage, it's the diagnosis: agents don't struggle with long documents because they can't retrieve facts, they struggle because they lose track of what they already found and why it mattered.
Call it a filing system for AI memory - unglamorous, but apparently the part everyone else forgot to build.