- CollectionAgent Harnesses & Operator Surfaces
- Classificationagent-native
- LicenseMIT
- MCPAvailable
- Agent SkillUnavailable
- InterfacesMCP · CLI
- Verified2026-09-26
What it does
LongHorizon-Harness wraps existing coding and computer-use agents in a repeated Manage–Execute–Audit loop. Every round rebuilds the next bounded step from the original goal, last verified checkpoint, failure evidence, and remaining work. The Executor starts fresh, the Auditor independently checks the real files/UI/logs/tests, and only accepted evidence becomes trusted progress. Run state survives context refreshes and failures.
Agent-facing interfaces
- Interaction pattern: CLI + agent adapters + optional computer-use plugin
- Catalog protocols: MCP, CLI
- MCP: ✅ Pass-through configuration
- Agent Skill: ⚠️ No standalone published Skill required
- Onboarding type:
cli
How agents onboard
Interaction pattern: CLI + agent adapters + optional computer-use plugin
# Python 3.10+; requires Codex or Claude Code on PATH
uv tool install lh-harness
cd /path/to/project
lh-harness init
lh-harness doctor
TASK="Inspect the current directory and summarize its files."
lh-harness run --task "${TASK}" --agent codex
See the full onboarding section in the dossier.
Protocol surface
| Interface | Detail |
|---|---|
| CLI | lh-harness init|doctor|run|web|dashboard|plugin |
| Configuration | Project .lh-harness/config.toml with per-role backend/model and limits |
| Agent adapters | Codex CLI, Claude Code, and custom AgentAdapter implementations |
| MCP | Native per-backend config passed into individual runs |
| Artifacts | Structured run state, event stream, role trajectories, audits, and reports |
Verification
Last verified 2026-09-26.
Sources:
- https://github.com/AMAP-ML/LongHorizon-Harness/releases/tag/v0.1.7
- https://lh-harness.pages.dev
- https://api.github.com/repos/AMAP-ML/LongHorizon-Harness