Same model, same backend. Swap only the loop wrapped around it, and WeaveBench completion goes 51.8 → 80.7.
LongHorizon-Harness wraps around Claude Code, Codex, OpenCode, and DeepSeek Harness. MIT licensed, 1,100 stars twenty days after launch. It doesn’t train the model and it doesn’t replace your agent. It only runs the loop.
Three roles: the Manager rebuilds the next step from the original goal plus verified progress plus evidence of failure. The Executor takes a fresh context and does only that step. The Auditor independently checks the real files and tests, and doesn’t take the Executor’s word for anything. Only what passes the check counts as progress; what gets bounced counts as evidence. If context is lost, you pick up from the last checkpoint. It turns acceptance into a role instead of one line of nagging in a prompt.
The Executor getting a fresh context on every step points in the same direction as the six rules of context engineering I wrote about earlier.
Their own reported numbers (Qwen 3.7-Plus plus Claude Code): OSWorld 2.0 full completion 2.8→8.3, Terminal-Bench 2.1 success rate 69.7→77.2 with 24% fewer tokens.