Skip to content
BranchesTags
 
 

Repository files navigation

A2A-T Multi-Agent Orchestration Center

Python Node.js License

A visual orchestration platform for multi-agent collaboration via the A2A-T protocol.
基于 A2A-T 协议的多智能体可视化编排平台。

中文

Overview

The Orchestration Center is a visual platform for designing and executing multi-agent workflows. It provides a drag-and-drop workflow designer, an async execution engine, and A2A-T negotiation integration — enabling teams to build, manage, and run complex agent collaboration flows without writing code.

Use cases: Telecom network assurance workflows, RAN energy-saving orchestration, SPN fault handling pipelines, enterprise multi-agent automation.

sequenceDiagram
    actor User as User
    participant FE as Workflow Designer<br/>(React :3003)
    participant BE as Backend :5001<br/>(FastAPI)
    participant LLM as LLM
    participant Reg as Agent Registry
    participant Agt as A2A Agents

    rect rgb(240, 248, 255)
        Note over User, Reg: 1. Agent Discovery
        FE->>+BE: GET /rest/v1/orchestrate/agent-cards
        BE->>Reg: Fetch AgentCards
        Reg-->>BE: AgentCard[]
        BE-->>-FE: agent-cards JSON
        FE-->>User: Display agent catalog
    end

    rect rgb(255, 250, 240)
        Note over User, Reg: 2. Workflow Creation (3 modes)
        alt 2a. PDF Import
            User->>FE: Upload PDF
            FE->>+BE: POST /rest/v1/orchestrate/parse-pdf
            BE->>LLM: Parse chapters & tasks
            LLM-->>BE: Structured preflow
            BE-->>-FE: PreFlow JSON
            FE->>BE: POST /rest/v1/orchestrate/generate-from-preflow
            BE->>LLM: Generate PSOP from PreFlow
            LLM-->>BE: PSOP workflow
            BE-->>FE: PSOP JSON
        else 2b. Manual Drag & Drop
            User->>FE: Drag agents, connect nodes, configure
            FE->>FE: Build workflow graph (React Flow)
        else 2c. Natural Language Intent
            User->>FE: Enter intent text
            FE->>+BE: POST /rest/v1/orchestrate/generate-from-intent
            BE->>Reg: Fetch AgentCards
            Reg-->>BE: AgentCard[]
            BE->>LLM: Generate PSOP from intent
            LLM-->>BE: PSOP workflow
            BE-->>-FE: PSOP JSON
        end
    end

    rect rgb(240, 255, 240)
        Note over User, Reg: 3. Save Workflow
        User->>FE: Click Save
        FE->>+BE: POST /rest/v1/orchestrate/workflows<br/>{psop: {...}}
        BE->>BE: Validate PSOP (Pydantic)
        BE->>BE: Persist (File JSON / PostgreSQL)
        BE-->>-FE: {workflow_id: "..."}
        FE-->>User: Saved successfully
    end

    rect rgb(255, 245, 255)
        Note over User, Agt: 4. Execute Workflow
        User->>FE: Click Execute
        FE->>+BE: GET /rest/v1/orchestrate/execute<br/>?psop_id=xxx&user_intent=...&lang=zh
        BE-->>FE: SSE: {"type":"init"}
        BE-->>FE: SSE: {"type":"start"}
        BE->>Reg: Fetch AgentCards for routing
        Reg-->>BE: AgentCard[]
        loop Per step (DAG traversal)
            BE->>BE: Build context from upstream outputs
            BE-->>FE: SSE: {"type":"agent_request",...}
            BE->>+Agt: A2A call (gRPC/HTTP)<br/>task + context
            Agt-->>-BE: Agent response
            BE-->>FE: SSE: {"type":"agent_response",...}
            opt A2A-T Negotiation
                BE->>Agt: Negotiation request
                Agt-->>BE: Negotiation response
                BE-->>FE: SSE: {"type":"negotiation_request",...}
                BE-->>FE: SSE: {"type":"negotiation_resolved",...}
            end
            opt Conditional Routing
                BE->>LLM: Route decision<br/>(JumpCondition matching)
                LLM-->>BE: Next step selection
            end
        end
        BE-->>FE: SSE: {"type":"psop_update",...}
        BE->>BE: Save ExecutionRecord
        BE-->>FE: SSE: {"type":"complete",...}
        BE-->>-FE: SSE: {"type":"close"}
        FE-->>User: Execution finished
    end
Loading

Features

Category Capability
Visual Designer React Flow-based drag-and-drop workflow builder with automatic Dagre layout
Multi-Mode Creation PDF document import, manual drag-and-drop, and natural-language-to-workflow via LLM
A2A-T Negotiation Fulfillment negotiation between agents via a2a-t-sdk, context carried in Task.metadata
Execution Engine DynamicWorkflowEngine — async DAG traversal, parallel A2A calls, conditional LLM routing, SSE streaming
Semantic Search Natural-language retrieval of previously built workflows
Dual API Layer Internal API (/rest/v1/orchestrate/*) for the frontend + External API (/api/v1/*) for third-party integration
SSE Streaming Real-time execution progress via 11 event types (init, start, agent_request, agent_response, psop_update, negotiation_request, negotiation_resolved, negotiation_failed, complete, error, close)
Pluggable Storage File-based JSON or PostgreSQL persistence via HandlerRegistry
Template Marketplace Pre-built workflow templates for telecom scenarios (live broadcast, energy saving, fault handling)
Sample Agents 11 sample A2A agents with negotiation support for testing and demonstration

Quick Start

Prerequisites

Component Requirement
Python 3.12+
Node.js 20.19+

Install & Run

# Clone the repository
git clone https://github.com/project-openan/orchestration-center.git
cd orchestration-center

# Backend setup
python3 -m venv .venv
source .venv/bin/activate      # Linux
# .venv\Scripts\activate       # Windows
pip install -r requirements.txt

# Start backend (port 5001)
python -m orchestrate.start

# Frontend setup (separate terminal)
cd workflow-designer
npm install --force
npm run dev                     # port 3003

# (Optional) Start sample agents
cd ..
python -m samples.start_agents_server

Verify

Service Check
Backend Uvicorn running on http://127.0.0.1:5001
Frontend Open http://localhost:3003 in browser
Sample Agents Agent startup messages in console

Architecture

flowchart TB
    subgraph frontend["Frontend"]
        wd["Workflow Designer<br/>React 18 + Vite + Tailwind<br/>Port 3003"]
    end

    subgraph backend["Orchestration Backend (Port 5001)"]
        direction TB
        api["Dual API Layer<br/>Internal /rest/v1/orchestrate/*<br/>External /api/v1/*"]
        domain["Core Domain<br/>PSOP Generator · Intent Generator<br/>Semantic Search · Publisher"]
        engine["DynamicWorkflowEngine<br/>DAG Traversal · Parallel A2A Calls<br/>LLM Routing · SSE Push"]
    end

    subgraph storage["Storage"]
        direction LR
        file[("File JSON")]
        pg[("PostgreSQL")]
    end

    subgraph agents["A2A Agents"]
        direction LR
        a1["Agent A"]
        a2["Agent B"]
        a3["Agent C..."]
    end

    wd -->|"REST / SSE"| api
    api --> domain
    domain --> engine
    engine --> file
    engine --> pg
    engine -->|"A2A Protocol<br/>+ A2A-T Negotiation"| a1
    engine --> a2
    engine --> a3

    style backend fill:#e1f5fe,stroke:#0288d1
    style frontend fill:#e8f5e9,stroke:#388e3c
    style storage fill:#f3e5f5,stroke:#7b1fa2
    style agents fill:#fff3e0,stroke:#f57c00
Loading

API Overview

External API (/api/v1/*)

Method Endpoint Description
POST /api/v1/orchestrate/sop SOP-based workflow orchestration (JSON text or file upload)
POST /api/v1/orchestrate/intent Intent-based workflow orchestration
GET /api/v1/orchestrate/psop/{id} Get PSOP workflow detail
POST /api/v1/orchestrate/search Search workflows by natural language intent
POST /api/v1/orchestrate/execute Auto-orchestrate + execute (SSE streaming)
GET /api/v1/orchestrate/execute/{id} Execute a known PSOP (SSE streaming)
GET /api/v1/executions List execution records
GET /api/v1/executions/{id} Get execution result

Internal API (/rest/v1/orchestrate/*)

Method Endpoint Description
GET /workflows List workflows
GET /workflows/{id} Get workflow detail
POST /workflows Create workflow
DELETE /workflows/{id} Delete workflow
POST /parse-pdf Parse PDF SolutionPackage and extract PreFlow
POST /generate-from-preflow Generate PSOP from PreFlow
POST /generate-from-intent Generate PSOP from intent
POST /retrieve-by-intent Retrieve workflow by intent
POST /retrieve-topn-by-intent Retrieve top-N workflows by intent
GET /agent-cards List available agent cards
GET /templates List workflow templates
POST /templates/{id}/import Import workflow from template
GET /execute Start workflow execution (SSE). Query params: psop_id, user_intent, lang
GET /execution-records List execution records
GET /execution-records/{id} Get execution record detail
DELETE /execution-records/{id} Delete execution record

Full API specification: API Reference

Configuration

Config File Purpose
etc/conf/server.conf Server IP, port, TLS certificates, persistence mode, registry URL
etc/conf/server.properties TLS versions, ciphers, rate limiting, connection limits
etc/conf/db_config.json PostgreSQL connection settings
common/config/llm_config.json LLM/embed/rerank model endpoints
common/config/README_en.md LLM configuration guide

A2A-T SDK Integration

This project integrates a2a-t-sdk for agent fulfillment negotiation:

A2AT_LLM_PROVIDER=deepseek
A2AT_LLM_MODEL=deepseek-chat
A2AT_LLM_API_KEY=<your-api-key>
A2AT_LLM_BASE_URL=https://api.deepseek.com
A2AT_NEGOTIATION_STATE_STORE_TYPE=in_memory

Configuration is auto-generated by common/a2at_config.py from common/config/llm_config.json.

Documentation

Document Description
User Guide Features, scenarios, quick start, FAQ
API Reference Full REST API specification
Developer Guide Custom handlers, LLM module, extension
GCP Deployment Guide Docker + GCP Cloud Run deployment guide
Frontend README Workflow Designer setup and tech stack
LLM Config LLM configuration reference

For Chinese documentation, see 中文 README or docs/zh/.

License

This project is licensed under the Apache License 2.0. See LICENSE for details.

About

No description, website, or topics provided.

Resources

Readme
License

Code of conduct

Code of conduct

Contributing

Contributing

Security policy

Security policy
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages