Long-Running Agents Need a Harness, Not a Bigger Context Window

What it does

LongHorizon-Harness is an orchestration layer that wraps existing agent backends—Claude Code, Codex, or anything with a lightweight AgentAdapter—to run multi-hour tasks across desktop apps and terminals without losing state. It breaks work into rounds where only independently verified results enter persistent task state, so a fresh context or a failed step doesn’t erase prior progress. Both GUI and CLI workflows are supported, often within the same continuous task.

The interesting bit

Rather than leaning on ever-larger context windows, the harness splits responsibility across three roles: a Manager keeps the goal and verified history, an Executor starts each round with clean context to tackle one defined task, and an Auditor inspects the real environment to decide what actually counts as done. This separation is the entire trick—no new model, just a stricter contract between planning, doing, and checking.

Key highlights

• Triple-role architecture: Manager, Executor, and Auditor can each use different models or backends to balance cost and quality.
• Verified state: only results that pass the Auditor’s inspection of files, interfaces, logs, and tests enter persistent task state.
• Fresh-context execution: the Executor starts each round clean, avoiding the bloat of an ever-growing conversation history.
• Cross-environment tasks: one workflow can move between browser, terminal, spreadsheet, and design software without losing the thread.
• Measurable gains: on the same Qwen 3.7-Plus backbone, WeaveBench pass rates rose from 51.8% to 80.7% and OSWorld 2.0 binary completion roughly tripled.

Caveats

• Windows support is included but has not yet been thoroughly tested; macOS is the current target.
• The project is iterating rapidly, and a new dashboard is noted as “in the works” with no committed date.

Verdict

Worth a look if you’re already running Claude Code or Codex and need agents to survive multi-hour, multi-app workflows without hand-holding. If your tasks finish in a single prompt, it’s probably overkill.