LongHorizon-Harness

Coding agents are good at bounded tasks but tend to drift, lose track of state, or simply run out of context on anything that takes hours rather than minutes, especially work that spans both GUI desktop apps and a terminal. LongHorizon-Harness, from Alibaba's AMap ML team, is built to fix that specific failure mode rather than to be another agent. It does not train a new model or replace Claude Code, Codex, OpenCode or DeepSeek Harness; it wraps whichever of those you already use in what the project calls a loop-engineering layer that keeps the agent oriented on one goal across a long-running desktop-and-CLI task.

The loop follows a fixed cycle: recover the goal and the last verified state, plan a single bounded next step, execute it with a fresh context window rather than an ever-growing conversation, check whether the real result on the desktop or in the terminal actually matches what was expected, and then either checkpoint that progress or feed the failure evidence into the next round. Because verification happens against the real computer state rather than the model's own claim of success, a step that silently failed does not get treated as done. The harness reports its results on named public benchmarks, including WeaveBench, OSWorld 2.0 and Terminal-Bench 2.1, and documents the approach in an accompanying arXiv paper.

This suits teams and individuals already using Claude Code, Codex, OpenCode or DeepSeek Harness for real work who want those same agents to keep running unattended for dozens of hours on tasks that mix desktop applications with command-line steps, such as a multi-stage data migration, a long research sweep, or an end-to-end app workflow that a single agent turn cannot finish. Because it is backend-agnostic, teams that already standardized on one coding agent do not have to switch tools to get the longer-running loop; they install LongHorizon-Harness on top of what they run today.

The project's initial burst was intense (eight releases and dozens of merged PRs between August 4 and August 20, 2026) but there has been no commit or release since August 20, over a month before this write-up, despite 40 open issues sitting unaddressed. The reported gains on WeaveBench, OSWorld 2.0 and Terminal-Bench 2.1 are the project's own results, not independently reproduced. Because LongHorizon-Harness only orchestrates an existing agent rather than replacing its judgment, the quality of any given step is still bounded by whichever backend model you point it at, and a systematically wrong verification step could let bad progress get checkpointed as good.

LongHorizon-Harness addresses a real and specific gap, agents that drift or lose state on long desktop-plus-CLI tasks, with a credible, benchmark-backed approach from a real engineering team rather than a solo side project. It is worth adopting if you already run Claude Code, Codex or a similar agent and need it to survive many hours of real work with checkpointed recovery. It is too early to lean on for anything mission-critical given the month-plus of silence since the last push; treat the initial three-week burst as promising and watch whether development resumes before depending on it in production.