title	Enterprise Workflow OpenEnv
emoji	🏢
colorFrom	blue
colorTo	indigo
sdk	docker
pinned	false

🏢 Enterprise Workflow OpenEnv

A real-world OpenEnv-compatible environment for training AI agents on enterprise procurement workflows — built for the Meta × PyTorch × Hugging Face OpenEnv Hackathon hosted by Scaler School of Technology (SST).

---

🚀 Overview

Enterprise Workflow OpenEnv simulates a realistic enterprise procurement pipeline where AI agents must complete multi-step workflows — from parsing purchase requisitions to drafting purchase orders and flagging approvals.

Designed to benchmark and train agents on structured decision-making in real-world business contexts.

---

🧩 Tasks

Difficulty	Description	Steps	Score
🟢 Easy	Parse purchase requisition → match correct inventory item	1	0.92
🟡 Medium	Parse requisition → check inventory → draft purchase order	3	0.92
🔴 Hard	Full pipeline: parse → inventory → supplier → PO → approval	5	0.92

---

🔌 API Endpoints

Method	Endpoint	Description
POST	/reset	Reset environment — works with or without body
POST	/step	Execute an agent action and get observation
GET	/state/{task_id}	Get current state for a task
GET	/tasks	List all tasks with action schemas
GET	/grader	Run all graders — returns sigmoid scores in (0,1)
GET	/baseline	Baseline scores for comparison

---

🏗️ Architecture

Environment Design

• Financial Guardrails — validates total_cost == quantity × unit_price on every PO draft
• Prerequisite Enforcement — agents cannot skip steps (no PO before inventory check)
• Graceful Error Handling — unknown actions return done=False allowing agent recovery
• Sigmoid Grading — smooth scoring strictly within (0, 1) using sigmoid normalization
• Partial Reward Signals — rewards at every step, not just episode end

Agent Design

• Skill System — composable, ordered workflow action definitions per task
• Episode Memory — remembers past workflow decisions across resets
• Jittered Backoff Retry — decorrelated exponential retry on LLM failures
• Trajectory Compressor — compresses long episode histories to fit token budgets
• Observation-Driven Loop — LLM decides next action based on environment observation

Action Space

parse_requisition → payload: {req_id, item_id} check_inventory → payload: {item_id} message_supplier → payload: {item_id} draft_po → payload: {item_id, quantity, total_cost, department} flag_approval → payload: {approver}

Observation Space

{
  "task_id": "easy|medium|hard",
  "step": 0,
  "result": {},
  "reward": 0.0,
  "done": false,
  "info": ""
}

---

🗂️ Project Structure

├── inference.py            # Main inference script (observation-driven LLM agent)
├── openenv.yaml            # OpenEnv configuration
├── Dockerfile              # Container definition (python:3.11-slim, port 7860)
├── requirements.txt        # Dependencies
├── server/
│   └── main.py             # Launch the server of uvicorn (app, host="0.0.0.0", port=8000)
├── app/
│   ├── main.py             # FastAPI routes and endpoints
│   ├── environment.py      # WorkflowEnvironment — state machine + reward logic
│   ├── models.py           # Pydantic typed models
│   ├── grader.py           # Sigmoid grading logic
│   └── mock_backend.py     # Mock enterprise backend (15 items, 13 requisitions)
├── agent/
│   ├── skills.py           # Composable skill system with episode memory
│   ├── retry_utils.py      # Jittered backoff retry
│   ├── trajectory.py       # Trajectory compressor
│   └── baseline.py         # Multi-episode baseline runner
├── tasks/
│   ├── easy.py             # Easy task definition
│   ├── medium.py           # Medium task definition
│   └── hard.py             # Hard task definition
└── tests/
    ├── debug_inference.py
    ├── final_validation.py # Full environment logic validation (5 tests)
    └── test_with_mock.py   # Mocked LLM tests

⚡ Quick Start

Reset Environment

curl -X POST https://ma4ku2-enterprise-workflow-env.hf.space/reset \
  -H "Content-Type: application/json" \
  -d '{"task_id": "hard"}'

Execute a Step

curl -X POST https://ma4ku2-enterprise-workflow-env.hf.space/step \
  -H "Content-Type: application/json" \
  -d '{"task_id": "easy", "action_type": "parse_requisition", "payload": {"req_id": "REQ-001", "item_id": "ITM-001"}}'

Check Grader Scores

curl https://ma4ku2-enterprise-workflow-env.hf.space/grader

Run Inference Agent

export OPENAI_API_KEY="<your_openrouter_key>"
export API_BASE_URL="https://openrouter.ai/api/v1"
export MODEL_NAME="openai/gpt-oss-120b:free"
python3 inference.py

🖥️ Self-Host Locally

# Clone the repo
git clone https://huggingface.co/spaces/MA4KU2/enterprise-workflow-env
cd enterprise-workflow-env

# Install dependencies
pip install -r requirements.txt

# Set your API key
export OPENROUTER_API_KEY="your-key-here"
export API_BASE_URL="https://openrouter.ai/api/v1"
export MODEL_NAME="openai/gpt-oss-120b:free"

# Launch the server
PYTHONPATH=$(pwd) python server/main.py
# → Server running at http://0.0.0.0:8000

Or use the convenience launcher:

./run.sh

---

📋 Baseline Scores

[START] task=easy env=enterprise-workflow-env model=openai/gpt-oss-20b:free
[STEP] step=1 action=parse_requisition reward=0.99 done=true error=null
[END] success=true steps=1 score=0.990 rewards=0.99
[START] task=medium env=enterprise-workflow-env model=openai/gpt-oss-20b:free
[STEP] step=1 action=parse_requisition reward=0.33 done=false error=null
[STEP] step=2 action=check_inventory reward=0.33 done=false error=null
[STEP] step=3 action=draft_po reward=0.33 done=true error=null
[END] success=true steps=3 score=0.990 rewards=0.33,0.33,0.33
[START] task=hard env=enterprise-workflow-env model=openai/gpt-oss-20b:free
[STEP] step=1 action=parse_requisition reward=0.20 done=false error=null
[STEP] step=2 action=check_inventory reward=0.20 done=false error=null
[STEP] step=3 action=message_supplier reward=0.20 done=false error=null
[STEP] step=4 action=draft_po reward=0.20 done=false error=null
[STEP] step=5 action=flag_approval reward=0.19 done=true error=null
[END] success=true steps=5 score=0.990 rewards=0.20,0.20,0.20,0.20,0.19

---

💻 Built On Constrained Hardware

Built entirely on a Raspberry Pi 5 (8GB RAM) running Kali Linux (aarch64/ARM64) OS — no cloud compute, no GPU.

All Docker builds, local testing, and HuggingFace deployments executed directly on ARM64 hardware, proving the environment is lightweight and portable.

---

🔑 Zero-Cost Inference Stack

Uses OpenRouter as API gateway with free-tier models — entire development pipeline completely free.

Compatible with any OpenAI-compatible endpoint via environment variables:

• OPENAI_API_KEY — your API key
• API_BASE_URL — swap to any provider
• MODEL_NAME — swap to any model

---

🗺️ Production Roadmap

Enterprise Integrations

• Temporal.io for durable long-running workflow execution
• Merge.dev for ERP integration (SAP, Oracle, Workday)
• Pinecone/Supabase for corporate procurement memory via RAG
• LangSmith for LLM observability and audit trails
• NeMo Guardrails for prompt injection protection
• CLI Interface to directly interact with the Agent Directly

Automation & Orchestration

• n8n for no-code workflow automation layer
• LangFlow for visual agent pipeline builder
• MCP servers for dynamic tool discovery and capability extension

Agent Infrastructure

• OpenClaw multi-agent orchestration for parallel procurement workflows
• Redis for short-term agent memory and API call caching
• Slack/Teams integration for Human-in-the-Loop approvals

---

🛠️ Built With

• FastAPI — High-performance API framework
• OpenEnv — Environment standard for AI agent training
• Docker — Containerized deployment
• Hugging Face Spaces — Deployment platform
• OpenRouter — Zero-cost LLM inference gateway

---

🏆 Hackathon

Submitted to the Meta PyTorch OpenEnv Hackathon, hosted by Scaler School of Technology (SST), Hugging Face, PyTorch, and Meta.

---

📄 License

MIT License

Made with ❤️ for the AI agent research community.