LongHorizon-Harness: The Loop Engineering That Lets Agents Run for Hours, Not Minutes

Every agent user hits the same wall: it can't go the distance. Give an agent a complex, multi-app task, and somewhere along the way it loses the plot — the context window fills up and it forgets its progress, or one step fails and the whole task restarts from zero, or it claims "done" and you have no way to verify. The problem isn't a lack of smarts. It's a missing layer of loop engineering: how an agent sustains state, verifies results, and recovers from failure over hours, not minutes. LongHorizon-Harness (1.5k stars, MIT) is built for exactly this — and it comes with an arXiv paper (2608.01964), which already tells you it's not another toy. The core loop Plan → act → verify → checkpoint or recover → repeat, until the work is actually done. Four mechanisms do the heavy lifting: Fresh-context execution — each step runs in a fresh context, so a full context window can't poison later steps. This is the direct cure for "the agent forgot what it was doing." Durable verified state — every step is verified; only verified progress is accepted. Checkpoint / recover — a failure doesn't restart the job; it resumes from the last checkpoint. Independent auditing — every step is auditable, so when something goes wrong you can find where. The README is refreshingly honest about what it is and isn't: it doesn't train new models or replace your agent — it wraps your agent in a persistent execution loop. Your Claude Code, Codex, OpenCode, or DeepSeek Harness stays itself, just with a shell that lets it run long and recover from failure. Two signals it's serious Paper + benchmarks. It ships with an arXiv paper and is measured on three public benchmarks — WeaveBench, OSWorld 2.0, Terminal-Bench 2.1. Hard numbers, reproducible evals. Rare in the agent space. It ships with an arXiv paper and is measured on three public benchmarks — WeaveBench, OSWorld 2.0, Terminal-Bench 2.1. Hard numbers, reproducible evals. Rare in the agent space. Model-agnostic. Four backends supported, no ecosystem lock-in. That's the mark of a pragmatic tool, not a vendor pitch. The honest boundary It improves persistence, not capability. If your agent can't operate a piece of software to begin with, the harness won't make it suddenly able to. What it gives you is stability over long horizons and recovery from failure — not more intelligence. It's also early (v0.1.x) and aimed at developers willing to run long tasks, not casual users. I've localized the README and core docs to Chinese: https://github.com/yangshun2005/LongHorizon-Harness-cn If you find this project useful, a star on the original repo supports the author's ongoing maintenance.