GitHub radar
LongHorizon-Harness: More Tasks Done by Claude Code
An execution and verification layer from the AMAP-ML team that wraps Claude Code and Codex, lifting task completion rates from ~50% to ~80% on benchmarks. Topped Hugging Face Daily Papers weekly ranking.
LongHorizon-Harness is an open-source execution, state-management, and verification system from the AMAP-ML team that runs on top of Claude Code and Codex CLI without replacing them. It splits work across three roles: a Manager that holds the original goal and verified progress, an Executor that starts each step with a fresh context, and an Auditor that independently checks results before they are recorded as done. On WeaveBench, task completion rises from roughly 50% to 80%; on OSWorld 2.0, full desktop task completion triples. The tool reached the top of Hugging Face Daily Papers weekly ranking on August 6, 2026.
Why a vibe-coder should care
If you regularly give Claude Code multi-step tasks it struggles to complete in one go — writing code, then running tests, then fixing errors, then deploying — LongHorizon-Harness gives the same model a structured scaffolding so complex work carries through to completion without losing track of what was already verified.
▌ More finds