Run long-horizon computer-use agent loops with LongHorizon-Harness
Use LongHorizon-Harness to keep Claude Code, Codex, DeepSeek Harness, or custom agent backends working through long GUI and CLI tasks with verified state, role separation, checkpoints, and recovery loops.
npx skills add agentskillexchange/skills --skill run-long-horizon-computer-use-agent-loops-with-longhorizon-harness
LongHorizon-Harness wraps existing agent runtimes in a loop-engineering workflow for long-running computer-use tasks. An operator gives the harness an outcome, then the system repeatedly plans the next bounded step, executes it in a desktop app or terminal, independently audits the real result, and checkpoints only verified progress before continuing.
What this skill actually does
Use this when a task is too long or fragile for a single agent session: cross-application desktop work, code-and-browser debugging, data processing followed by visual validation, or multi-round operations where context refresh, failure recovery, and audit evidence matter. The workflow can run from the browser workbench with `lh-harness web –workspace-root .` or from the CLI with `lh-harness run –task … –agent codex`.
The scope is the repeatable operator loop around existing agents, not a new model, generic SDK, benchmark paper, or broad automation platform. The approved skill is bounded to installing the `lh-harness` command, initializing project configuration, selecting supported agent backends and computer-use plugins, running tasks, and inspecting the harness reports/checkpoints that prove progress.