AgentIR

Agent Workflow Intermediate Representation — a framework-agnostic bridge between Dynamic Workflow (AI-generated) and Traditional Workflow (BPMN).

English | 中文

---

Introduction

AgentIR solves a fundamental contradiction:

• Traditional workflows (BPMN drag-and-drop) can build complex logic but require deep expertise
• Dynamic Workflows (AI writes scripts from natural language) have zero barrier but lack visualization, testability, and clear node definitions

AgentIR bridges the gap: describe workflows in natural language → generate AgentIR (JSON IR) → visually review/edit → compile to multi-framework executable scripts.

Natural Language → Planner (LLM) → AgentIR (.json)

                                        ├── Visual Editor (ReactFlow)
                                        ├── Security Audit
                                        ├── ADK Compiler
                                        ├── LangGraph Compiler
                                        └── CrewAI Compiler

Current capabilities:

• 7 workflow node types (agent, tool, subworkflow, sequence, parallel, condition, loop)
• AI auto-generates system prompts for each Agent
• ReactFlow visual editor (property panel, undo/redo, IR diff, dagre layout)
• 3 compiler targets (Google ADK, LangGraph, CrewAI)
• 289 tests, 100% pass rate

---

Screenshots

📸 （Screenshots to be added: workflow generation, visual editor, execution results）

---

Getting Started

Prerequisites

• Python 3.12+
• Node.js 18+
• DeepSeek API Key (or any OpenAI-compatible API)

1. Clone

git clone <repo-url>
cd DynamicWorkflowAST

2. Configure API Key

cp .env.example .env

Edit .env with your DeepSeek API key:

AGENTIR_LLM_PROVIDER=deepseek
AGENTIR_LLM_MODEL=deepseek-chat
AGENTIR_LLM_API_KEY=sk-your-key-here

3. Create virtual environment & install dependencies

python3 -m venv .venv
source .venv/bin/activate
pip install openai fastapi uvicorn python-dotenv pytest

4. Start backend

agentir-server
# or
python -m agentir.server.cli

Open http://localhost:8000/docs for the API documentation.

5. Start frontend (optional)

cd ui
npm install
npm run dev

Open http://localhost:5173 for the visual editor.

6. Quick test

# Generate a workflow
curl -X POST http://localhost:8000/api/v1/workflows/generate \
  -H "Content-Type: application/json" \
  -d '{"requirement": "First have researcher gather data, then writer produce a report"}'

# Run all tests
pytest