# AMAP-ML/LongHorizon-Harness

The long-horizon computer-use harness. Run AI agents across desktop apps and the CLI for extended periods while preserving task state and making reliable progress on complex workflows. Features fresh-context execution, durable verified state, independent auditing, recoverable progress, and native Claude Code / Codex / OpenClaw integration.

Repository: https://github.com/AMAP-ML/LongHorizon-Harness
Canonical: https://ross.abutalabs.com/products/longhorizon-harness
Homepage: https://lh-harness.pages.dev
Language: Python
License: MIT
License Family: permissive
Topics: agent, harness, long-horizon, long-horizon-agents, longhorizon-harness, claude, claude-code, claude-plugin, cli, codex, codex-desktop, codex-plugin, cua, gui, loop, loop-engineering, deepseek-harness
Last push: 2026-08-20T12:36:03+00:00

## Health v2 (maintenance only)
Score: 79/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 98, release rhythm 98, longevity 2
- inputs: {"age_days": 29, "days_push": 13, "days_rel": 13, "gap_med": 2, "n_releases_24m": 8}
- flags: young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1332, forks 136 (observed 2026-08-28T04:04:23.909528+00:00)

## What it is
LongHorizon-Harness is a Python-based loop engineering harness that turns existing coding agents (Claude Code, Codex, OpenCode, DeepSeek) into long-running computer-use systems across desktop apps and the CLI. It manages a plan-act-verify-checkpoint loop with durable, independently audited task state so agents can make reliable progress on multi-hour tasks.

## Use cases
- run AI agents on long-horizon tasks for dozens of hours
- keep an agent working across desktop apps and the terminal
- recover agent progress after failure or context refresh
- independently verify each step an agent completes
- benchmark agents on WeaveBench, OSWorld, and Terminal-Bench
- orchestrate different models for manage, execute, and audit roles
- checkpoint and resume complex multi-step workflows

## When to choose
- you need an existing agent like Claude Code or Codex to run reliably for extended periods
- your tasks span GUI desktop apps and CLI tools
- you want audited, verifiable progress records outside the agent session
- you want to improve benchmark scores without changing the underlying model

## When to avoid
- you need a simple single-shot agent run without state management
- you want to train or fine-tune a model rather than orchestrate one
- you need a fully managed cloud service rather than a local Python harness
- your agent runtime is not supported (no Claude Code, Codex, OpenCode, or DeepSeek backend)

## Facets
- artifact type: framework
- maturity: active
- function: agent-framework, workflow-automation, developer-tools, cli, gui
- domain: developer-tools, large-language-models
- platform: windows, cli, python
- tags: loop-engineering, computer-use, long-horizon-agents, claude-code, codex, agent-harness, checkpointing, task-orchestration, mea-loop, ai-agents, automation, linux, macos

## Member repositories
- AMAP-ML/LongHorizon-Harness (main) score 79

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:23.909528+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:44:56.121163+00:00, confidence not recorded.
  - readme: https://github.com/AMAP-ML/LongHorizon-Harness (fetched 2026-08-28T04:04:23.909528+00:00, sha 75b0aa90cb69)
  - homepage: https://lh-harness.pages.dev (fetched 2026-08-29T12:04:13.921941+00:00, sha e7b2264fdf9e)
- Data as of 2026-08-30T08:39:29.467469+00:00.