Alibaba's Amap Lab Published a Long-Horizon Computer-Use Harness With an arXiv Paper and Three Benchmarks

GitHub·medium signalXBlueskyLinkedInCopy link

`AMAP-ML/LongHorizon-Harness` (675 stars since 2026-08-04, Python, MIT, arXiv 2608.01964) wraps Claude Code or Codex in a plan → act → verify → checkpoint-or-recover loop meant to hold a single goal across desktop apps and the CLI for dozens of hours. Each round executes a bounded step with a fresh context, checks the actual result in the real computer, then either checkpoints accepted progress or feeds failure evidence forward — no new model, no agent replacement, just the durable loop around one. Results are claimed on WeaveBench, OSWorld 2.0 and Terminal-Bench 2.1, and it made Hugging Face Daily Papers; the framing that the model decides one round and the harness decides the loop is the argument to evaluate.