Agentic Business Process Modeling

A research prototype for AI-assisted BPMN process modeling. The system employs a single, cohesive autonomous AI agent that generates, refines, and validates BPMN 2.0 diagrams directly in the browser, guided by natural language task descriptions and iterative goal-based planning.

Note: This tool was developed as part of academic research. See the Citation section at the bottom of this document.

---

Table of Contents

• Features
• Architecture Overview
• Requirements
• Setup
• Running the Application
• Usage
• Project Structure
• Citation

---

Features

• Autonomous Agentic Orchestration: A single AI agent handles planning, executes modeling actions on the BPMN canvas, and performs internal self-reflection to validate outputs.
• Self-Reflective Iteration: The agent defines sub-goals, drafts the model, and reviews the draft against semantic and syntactic rules. The agent iterates internally up to three times to resolve detected issues before presenting the final result.
• Human-in-the-Loop Escalation: If the agent cannot resolve all issues within the internal loop, the system halts autonomous execution and hands control back to the user for targeted feedback.
• Live BPMN Canvas: Full bpmn.io modeler running in the browser — the agent creates, updates, and deletes elements in real time.
• Visual Feedback Markers: Structural validation of the generated diagram surfaces directly on the canvas using severity-graded visual markers (Info / Warning / Critical).
• Custom Tasks: Upload your own process description (.txt, .md, .pdf, .docx) or type instructions directly.
• Conversational Q&A: Ask the agent conceptual questions about BPMN or the current model without triggering a full re-modeling cycle.
• Export: Download the current diagram as a .bpmn file at any time.

---

Architecture Overview

User ──► Autonomous Agent (PLAN → EXECUTE → SELF-REVIEW)
              │
              ▼
         BPMN Canvas (bpmn.io)

All LLM input/output uses our custom serialization format. Designed to be highly compact and LLM-friendly, the notation drastically reduces data usage by 30 to 80 percent compared to conventional JSON files, minimizing context window overhead and latency.

---

Requirements

• Python 3.10 or higher
• An OpenAI API key

---

Setup

1. Clone the repository

git clone [https://github.com/](https://github.com/)<your-username>/<your-repo>.git
cd <your-repo>

2. Create a virtual environment

python -m venv venv

Activate the virtual environment:

OS	Command
Windows	venv\Scripts\activate
macOS / Linux	source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Configure environment variables

Create a .env file in the project root:

OPENAI_API_KEY=sk-...
GPT_MODEL=gpt-5.2
SECRET_KEY=your-secret-key-here
FLASK_DEBUG=false

Variable	Required	Description
OPENAI_API_KEY	✅	Your OpenAI API key
GPT_MODEL	✅	Model identifier (for example gpt-5.2 or gpt-4.1)
SECRET_KEY	✅	Flask session secret — use any long random string
FLASK_DEBUG	❌	Set to true for development hot-reload (default: false)

---

Running the Application

python main.py

The server starts at http://127.0.0.1:8080.

Open your browser and navigate to that address to access the task selection screen.

---

Usage

1. Select or create a task

On the home screen you will see a list of predefined modeling tasks. Click Start on any task to open the modeling canvas, or use Custom Task to supply your own process description.

For custom tasks you can:

• Type or paste a process description directly into the text field.
• Upload a .txt, .md, .pdf, or .docx file — the application extracts the text automatically.

2. Interact with the AI agent

Once a task is open, the right-hand panel contains the chat interface. The system starts automatically:

• The Agent reads your task and defines a set of modeling goals.
• The Agent executes the goals step by step on the BPMN canvas (you can watch the diagram being built live).
• Upon finishing the draft, the Agent conducts an internal self-reflection to detect and fix errors.

You can interact at any point:

What you want	How
Ask a question about BPMN or the diagram	Type and send your message — the agent answers without re-modeling
Request a change	Describe the change — a new planning cycle begins
Stop the current iteration	Click Stop

3. Review the plan panel

The left sidebar shows the current modeling goals and the respective status (pending / in progress / complete). This panel updates in real time as the agent works.

4. Inspect validation issues

After the self-reflection phase, any detected issues appear in the Review panel below the canvas with severity levels:

Severity	Meaning
🔴 Critical	Structural error or semantic deviation preventing valid BPMN
🟡 Warning	Suboptimal modeling structure needing attention
🔵 Info	Non-mandatory suggestion or best-practice note

5. Complete the task

Click Confirm Solution when you are satisfied with the diagram.

• If only informational issues exist (or no issues at all), the task completes immediately and you are returned to the home screen.
• If warnings or critical issues were flagged in the last review, the system blocks the thoughtless completion of the task. A confirmation dialog asks whether you explicitly grant permission to ignore the highlighted issues. You can Cancel to continue working or Confirm to finish regardless.

6. Export

Click the Export button at any time to download the current BPMN diagram as an XML file compatible with any BPMN 2.0-compliant tool (for example Camunda Modeler, Signavio, or bpmn.io).

---

Project Structure

.
├── main.py
├── config.py
├── requirements.txt
├── .env                     # (not committed) API keys & secrets
│
├── app/
│   ├── ai_service.py
│   ├── prompts.py
│   ├── sockets/
│   │   └── chat_handler.py
│   ├── static/
│   │   ├── css/
│   │   └── js/task.js
│   └── templates/
│       ├── index.html
│       └── task.html
│
├── lion/
│   ├── encoder.py
│   └── decoder.py
│
├── utils/
│   └── bpmn_validator.py
│
└── benchmarks/
    ├── runner.py
    ├── bpmn_executor.py
    └── results/

---

Citation

If you use this tool or build upon the code in your research, please cite the following:

Will be updated for the BPM Proceedings 2026

@inproceedings{anonymous2026agentic,
  title     = {[Paper Title Placeholder]},
  author    = {[Author(s) Placeholder]},
  booktitle = {[Conference/Workshop Name Placeholder]},
  year      = {2026},
  note      = {[Additional details placeholder]}
}