README.md

metadata

license: mit
language:
  - en
size_categories:
  - n<1K
task_categories:
  - other
tags:
  - benchmark
  - computer-use-agent
  - gui
  - cli
  - hybrid-interface
  - long-horizon
pretty_name: WeaveBench

WeaveBench

A long-horizon, real-world benchmark for computer-use agents with hybrid GUI + CLI + code interfaces.

📄 Paper: github.com/Wanli-Lee/WeaveBench (paper.pdf) 💻 Code: github.com/Wanli-Lee/WeaveBench

WeaveBench evaluates whether a single agent can orchestrate visual desktop control, command-line execution, code editing, browsers, and external tools within one long-horizon workflow. Best observed pairing in the paper: Claude Opus 4.7 + Claude Code at 41.2 % PassRate — far from saturation.

This HF repository hosts:

1. tasks/ — the 114 paper-final tasks across 8 domains, ready to drop into the WeaveBench orchestrator.
2. runtime_assets/ — pre-built bootstrap tarballs for the 4 supported in-VM agent harnesses (OpenClaw, Codex CLI, Claude Code, Hermes).

---

1. Layout

WeaveBench/
├── tasks/
│   ├── batch1/    batch2/    batch3/    batch_gen/   ← 4 release batches
│   │   ├── DAV/   DES/  DOC/  DSK/  GAM/  OPS/  SPA/  WEB/    ← 8 domains
│   │   │   └── <DOMAIN>_task_<NN>_<slug>.md          ← one file per task
│   │   └── workspace/<DOMAIN>/<task_dir>/exec/       ← per-task scaffold
│
└── runtime_assets/
    ├── openclaw.tar.gz                 491 MB   (reference harness)
    ├── codex.tar.gz                    125 MB   (OpenAI Codex CLI)
    ├── claudecode.tar.gz                69 MB   (Anthropic Claude Code)
    ├── hermes.tar.gz                   121 MB   (Nous Research Hermes)
    └── hermes_mcp_wheels.tar.gz          9 MB   (offline mcp wheels for Hermes)

Total: 114 tasks across 8 domains, ~203 MB of task content + ~815 MB of runtime tarballs.

2. Quick download

# Install the code package first:
pip install git+https://github.com/Wanli-Lee/WeaveBench.git

# Then fetch dataset + runtimes:
weavebench-download-dataset --dest ./cache         # tasks/
weavebench-download-assets  --dest ./cache         # runtime_assets/

# Run one task end-to-end with OpenRouter:
export OPENROUTER_API_KEY=sk-or-v1-...
weavebench-run \
    --harness openclaw --transport messages \
    --model anthropic/claude-opus-4 \
    --tasks_root ./cache/tasks \
    --bench_subdirs batch_gen --categories WEB --task_filter task_1 \
    --result_dir ./results/smoke

Or download manually from this repo via the HuggingFace web UI.

3. Per-domain task counts

Domain	Description	Count
WEB	Web inspection / DevTools / browser auditing	12
DAV	Data-analyst / SRE workflows (Jaeger, Streamlit, dbt, …)	13
DOC	Document / slide / brainstorm editing	14
DSK	Desktop / system / IDE tooling	15
GAM	Gamedev / Godot scene editing	17
OPS	Web-ops, nginx, dashboards	12
SPA	SPA / front-end stateful flows	11
DES	Inkscape / vector / design	~10
Total		114

4. Task .md schema

Each task file has these sections in order:

# <Human title>

## Goal             — short user request (what the agent reads as `instruction`)
## Setup            — preconditions assumed to be true in the VM
## Warmup           — bash commands the orchestrator runs before the agent starts
## Expected Output  — files the agent must produce in /tmp_workspace/results/
## Grader           — Python `def grade(workspace_path, transcript) -> dict`

The grader returns {"score": float ∈ [0, 1], "scores": {sub_rubric: float, ...}, "msg": "..."}. The grader never sees the chat transcript (transcript=[]), so it must be reproducible from artifacts alone. The paper's trajectory-aware judge (weavebench/eval/agent_judge) is the second-pass defense against fabricated visual evidence.

5. Citation

@article{li2026weavebench,
  title  = {WeaveBench: A Long-Horizon, Real-World Benchmark for Computer-Use Agents with Hybrid Interfaces},
  author = {Li, Wanli and Zhou, Bowen and Yang, Yifan and Yu, Yunyao and Li, Dongsheng and Xu, Zhou and Shan, Caihua},
  year   = {2026},
  month  = {May},
}

6. License

• Tasks: MIT.
• Runtime tarballs: each tarball repackages third-party software (Codex CLI is Apache-2.0; Claude Code, Hermes, OpenClaw retain their upstream terms). See NOTICE in the code repo for full attribution.