FORTE: Full-Cycle Office Real-World Task Evaluation

A general agent benchmark on daily office productivity across 15 corporate professions, including multi-modal office environments (xlsx / docx / pptx / pdf / csv / images / source files).

GitHub

What's in the benchmark

Built to be running on a single Docker image: agent + judge + LibreOffice + CJK fonts all live in one container.

Real Office Artifacts

xlsx, docx, pptx, pdf, csv, images, source files.

15 Corporate Professions

Marketing, Sales, BA, Operations, Dev, SRE, HR, Finance, PM, Legal, Algorithm, QA, UI/UX, Administration, and General.

LLM-Based Rubric Grading

Output artifacts are checked against expert-annotated rubrics by an LLM judge.

Leaderboard

All models are evaluated with thinking mode enabled. For models that expose a configurable reasoning effort / thinking level, the highest setting is used.

# Model Avg@3 Pass@3 Pass^3
loading…

Quickstart

A full run is three commands once your gateway is configured.

# 1. Configure an OpenAI-compatible gateway (used for both agent and judge).
cp openclaw_config/openclaw.json.example openclaw_config/openclaw.json
cp openclaw_config/.env.example          openclaw_config/.env

# 2. Build the single agent+judge Docker image.
bash docker/fetch-lo.sh
docker build -f docker/Dockerfile.agent -t forte-agent:latest docker/

# 3. Run.
bash scripts/run.sh

# Or run the benchmark CLI directly with an explicit model.
python scripts/benchmark.py \
  --model your-gateway/claude-sonnet-4.6 \
  --dataset ./data \
  --runs 3 \
  --concurrency 1 \
  --output-dir ./results

# Optional overrides.
MODEL=your-gateway/claude-sonnet-4.6 RUNS=3 CONCURRENCY=2 bash scripts/run.sh

Resources

GitHub Repository →

Task Definition, Source Code, Docker Recipe, Judge Implementation.