↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Harnesses/

LongHorizon-Harness

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~980 words)
agent-curated 533 computer-use 2 fully-ai-generated 636 harness 7 llm=glm-5.3-flash 544 long-horizon 2 meta-harness 3 research-note 503
Table of Contents

LongHorizon-Harness is MIT-licensed research code from Alibaba’s DreamX team (the Amap ML organization) that turns existing agents into long-running computer-use systems through a Manage-Execute-Audit loop: a manager plans one bounded subtask at a time, a fresh-context executor performs it, and a read-only auditor certifies what actually changed before anything counts as progress.

It is the meta-harness with an audit gate: four members of this very category (Claude Code, Codex, OpenCode, DeepSeek Harness) serve as its interchangeable backends, and only auditor-verified facts may enter the task state that survives between rounds.

What it is
#

A Python package (uv tool install lh-harness) that wraps existing agent CLIs through a lightweight AgentAdapter, preserving each backend’s native loop while the harness controls context, permissions, budgets, and reporting. The three roles can each use a different model or backend by configuration alone, with per-round budgets (1,800 seconds for the executor, 300 for manager and auditor) and a 25-round cap per task. Execution spans desktop GUI (click, type, scroll) and terminal (code, commands, dependencies) in one continuous task, with a React/FastAPI web dashboard (lh-harness web) and per-run audit trails under runs/. The accompanying paper (arXiv 2608.01964, submitted 2026-08-03) formalizes the loop as independently audited state transitions and reached #1 on the Hugging Face Daily Papers weekly ranking for 2026-W32.

Status
#

Published and then quiet: 1,729 stars, 188 forks, and 43 open issues as of 2026-10-13, created 2026-08-04, with the last push and latest release v0.1.7 both on 2026-08-20, an eight-week silence as of this check. The release train ran daily-to-weekly through August (v0.1.4 dashboard, v0.1.5 DeepSeek Harness support, v0.1.6 OpenCode support, v0.1.7 conversational follow-ups) and has not moved since. Self-reported results: on WeaveBench, Qwen 3.7-Plus through Claude Code moves from 51.8 to 80.7 percent PassRate; Terminal-Bench 2.1 from 69.7 to 77.2 percent with 24 percent fewer tokens; OSWorld 2.0 binary completion from 2.8 to 8.3 percent, with gains carrying to Claude Opus 4.7. PyPI downloads run about 1,000 per month as of 2026-10-13, and independent discussion is one 1-point Hacker News story.

Star History Chart

Strengths
#

  • The audit gate is a real design contribution: the read-only auditor never sees the executor’s reasoning, its reports are the only cross-round memory, and detected mutation is recorded as an integrity violation that cannot support completion.
  • Fresh-context execution attacks context rot directly: each round starts clean, and only compact audit reports persist, so long-run degradation is bounded by design rather than by prompt tricks.
  • Backend independence is genuine and current: Claude Code, Codex CLI, OpenCode, and DeepSeek Harness are all wired, each role assignable to a different model.
  • The evaluation is unusually self-critical for a vendor report: the paper discloses that its WeaveBench runs use root privileges while official baselines use a regular user, and reports where the harness costs more tokens than it saves.
  • Corporate provenance with a paper trail: an arXiv manuscript, a #1 weekly papers ranking, and a named team give the eight-week-old project unusual provenance for its size.

Cautions
#

  • Eight weeks without a commit or release as of 2026-10-13: active enough to judge, too quiet to bet a workflow on, and the archive-or-revive question is live if the silence extends.
  • The headline WeaveBench gain is against same-backbone baselines; the comparison with official leaderboard numbers crosses a privilege setting the paper itself flags as unmatched.
  • Absolute computer-use numbers stay low (8.3 percent binary completion on OSWorld 2.0), so the loop multiplies a backbone’s capability rather than replacing it.
  • Token cost is backbone-dependent: 2.3 times baseline on WeaveBench and 3.6 times on OSWorld 2.0 output for Qwen 3.7-Plus, against a 24 percent saving on Terminal-Bench 2.1.
  • Provenance concentration: one corporate lab, one paper, no external reproduction, and a community footprint of one 1-point Hacker News story.

Pricing
#

Free and open source under MIT. No hosted tier; cost is your own model bills, with the paper’s token-share analysis (manager about 2 to 8 percent, auditor 19 to 38 percent) as the published cost expectation.

Compared to
#

  • Omnigent: the other meta-harness, wrapping the same class of CLIs behind one API for governance and session sharing; Omnigent adds policy and sandboxes, LongHorizon-Harness adds the audited long-horizon loop.
  • Claude Code: the wrapped incumbent; use it directly for interactive work, and LongHorizon-Harness when tasks outlast context windows and unverified self-assessment is the failure mode.
  • Mini-SWE-agent: the minimal-loop contrast; Mini-SWE-agent is the baseline this category measures against, and LongHorizon-Harness is the argument that the loop, not the model, is what the baseline lacks.

Bottom line
#

Recommended for research and evaluation workloads with hours-long computer-use tasks, a strong wrapped backend, and tolerance for a self-reported evidence base. Not for production daily driving while the project stays silent, and not for tasks whose bottleneck is model capability the loop cannot add.

Changes
#

  • 2026-10-13 - Created.

See also
#

References
#