LongHorizon-Harness

Adopt it if you already build long-running agents and your main failure mode is “the model says it finished, but the world state is wrong.” If you just want a ready-made general AI assistant, this is not that. More accurately, it is not a new foundation model or a full SaaS workflow product; it is closer to a state-management and independent-validation runtime layered on top of executor-style agents such as Claude Code or Codex.

The available evidence consistently describes a split between Manager, Executor, and Auditor. The executor acts, then an auditor checks real environment state before the next step is accepted. Several posts repeat benchmark claims that swapping only the harness improved results on WeaveBench, OSWorld 2.0, and Terminal-Bench 2.1, with some token savings. But popularity proof and usability proof should be separated: reposts and “top paper” mentions mainly show attention, while the stronger evidence comes from posts describing independent verification, error-injection style checks, and same-model control comparisons. Even so, most evidence here is still social-media retelling, with limited official repo material, tutorials, or broad third-party reproductions.