A BPMN-based Multi-Agent Orchestration Framework for the Enterprise

Process GPT is an open-source platform that combines the 30-year-proven BPMN (Business Process Model and Notation) international standard with the autonomy of modern AI agents. It lets non-technical users design business processes in natural language, lets multiple specialized agents collaborate to execute them, and continuously learns from feedback to make every future run better.

Where classic BPM requires code and specialists, and pure agent frameworks (CrewAI, LangGraph, AutoGen, Swarm/OpenAI SDK) leave you without visual processes, audit trails, or compensation semantics — Process GPT sits at the intersection: visual + code + standards-based, engineered for production.

Online Service 🌐 www.process-gpt.io

대화로 HWPX 문서를 생성하고 편집하는 ProcessGPT	HWPX 문서를 생성하는 AI ProcessGPT	나노바나나 기반 슬라이드 생성 ProcessGPT
6. ProcessGPT 아키텍처	ProcessGPT 26년 1월 웨비나_빈 캔버스 위, AI가 그리는 비즈니스의 미래 (Full)	Agent Ops 구현 미리보기 - [Process GPT] 휴먼-에이전트 성과 분석
ProcessGPT 웨비나 Full	ProcessGPT AI 에이전트 팀이 만드는 완전 자율 업무 시대	Process GPT - 나의 친구, 나의 AI 에이전트 (full)
Process GPT - 나의 친구, 나의 AI 에이전트	[Process-GPT] AI를 이용한 비즈니스 프로세스 / 폼 자동 생성	ProcessGPT 1차 프로모션 영상
Process-GPT e2e 데모	process-gpt 채팅 기능 데모	process-gpt 인스턴스 실행 데모

• Flexible and Robust Collaboration via Multi-Agent Systems — Process GPT configures multiple AI agents to collaborate within a single BPMN-based business process, so that multiple agent frameworks can professionally handle complex tasks and share intermediate results. This reliably automates high-complexity work that would be difficult for a single agent. Each agent leverages specialized domain knowledge and tools, and can automatically call upon other specialized agents to delegate tasks when necessary.

• Automated Business Process Generation — Process GPT is designed to let AI agents automatically define business processes, producing results without constant human instruction or manual execution of every step. This minimizes human intervention, embodying the ambient agent philosophy, and ensures that human involvement does not become a bottleneck.

• Natural-Language-Based Continuous Process Learning and Optimization — Non-expert users can define business processes using natural language, which automatically generates initial process models. An automatic optimization cycle analyzes user feedback and agent execution logs, continuously improving processes and augmenting training data for workflows that need correction.

• Deterministic Regularization of AI Decisions — When agents repeatedly make the same kind of judgment, Process GPT automatically converts it into a DMN decision table or Python rule, guaranteeing that "same input = same output" for enterprise-critical paths while keeping AI flexibility for exceptions.

• Enterprise-Grade Reliability — BPMN Compensation Events provide automatic rollback and compensating transactions on failure. Human-in-the-Loop is a native BPMN pattern, not a workaround. Every step is auditable against an ISO/IEC 19510 process model.

• Framework-agnostic Multi-Agent System — powered by LangChain Deepagents, CrewAI, and more; pick the best runtime per task
• BPMN-based Hybrid Process Execution — deterministic (DMN/code) and stochastic (LLM reasoning) modes in one diagram
• Collaborative Work via the Agent-to-Agent (A2A) Protocol — agents discover and negotiate with each other through Agent Cards
• Isolated Agent/Tool Execution — each MCP and A2A server runs in its own container, orchestrated on Kubernetes with KEDA queue-based autoscaling and a Sidecar isolation pattern
• Skill Self-Learning & Feedback Loops — agents improve their own Skills through a Think → Execute → Reflect cycle driven by user feedback
• Context Engineering — Mem0 + Neo4j knowledge graph + Memento RAG service give agents deep organizational context
• Voice & Realtime Channels — GPT-4 Realtime API + Twilio PSTN integration for voice-driven process triggers
• Process Marketplace — share and reuse verified process templates across teams and organizations
• Integrations — Browser-use, OpenAI Deep Research, Supabase (Postgres, Realtime, Storage, Auth), ERP/CRM via MCP, N8n (coming soon)

Process GPT is the only player in the "BPMN + AI Hybrid" category — purpose-built for enterprises that need the governance of BPM and the autonomy of modern agents at the same time.

• Enterprises modernizing legacy BPM — keep your BPMN assets, add AI autonomy
• Regulated industries — finance, healthcare, public sector where audit trails and compliance are non-negotiable
• Citizen developers — business users automating their own work in natural language, no coding required
• AI teams building production agents — skip the infrastructure rebuild; get K8s-native isolation, scaling, and observability out of the box

• Website: process-gpt.io
• Documentation: docs.process-gpt.io
• SaaS: try it instantly at process-gpt.io
• Self-host: docker compose up deploys the full stack; Kubernetes manifests included for production

Process GPT creates a new category — "the BPMN of AI agents" — and sets the standard for enterprise multi-agent orchestration.

Maintained by uEngine Solutions · learning@uengine.org

• execution (Execution Engine): GitHub
• memento (Document Memory Storage): GitHub
• crewai-action (MCP / Multi-Agent Task Execution Agent): GitHub
• crewai-deep-research (Multi-Agent Deep Research Agent): GitHub
• openai-deep-research (OpenAI-based Deep Research Agent): GitHub
• react-voice-agent (Voice Interaction Agent): GitHub
• API gateway: GitHub
• frontend: GitHub
• completion: GitHub
• autonomous-execution: GitHub
• agents.github.io: GitHub
• generic-agent: GitHub
• agent-feedback: GitHub
• mcp-validator: GitHub
• agent-sdk: GitHub
• langchain-react: GitHub
• a2a-orch: GitHub
• agent-utils: GitHub
• bpmn-extractor (ProcessGPT BPMN extractor from PDFs): GitHub
• computer-use: GitHub
• claude-skills (MCP server for searching and retrieving Claude Agent Skills using vector search): GitHub
• deep-research: GitHub
• office-mcp: GitHub

• Robo Architect: GitHub

Users should be able to declare and modify processes, rules, system integration mechanisms, etc. in natural language, and the system should automatically improve with minimal feedback provided during use.

All such changes must be logged for tracking and recovery, while users should simultaneously be able to directly control automation results and regulations through a generalized UI at any time.

• All process definitions, rules, system integrations, and business interfaces should be writable in natural language without requiring programming knowledge or complex logical/mathematical thinking.
• Users should be able to design automation with business objective or strategic-level descriptions alone, without undergoing separate training processes.
• The system should be progressively refined and managed through minimal feedback (approval, modification, rejection) provided during actual use.

• Automated agents must provide human interfaces that allow people to substitute and perform tasks at any time.
• Each task should provide necessary context (related data, previous step outputs, similar cases) in a clear and organized manner to facilitate human processing.
• Agents learn from actual performance examples where humans directly handle tasks, correcting and improving their execution knowledge. In other words, human exemplars become the agent's training data.

• When errors or failures occur in automated processes performed by agents, recovery should be automatically implemented through compensating transactions (rollback).
• Operators should not need to track and correct agent details individually; the system itself should take responsibility for failure recovery and processing.
• This liberates users from system imperfections and ensures overall business continuity.

Please refer to the Local Installation Guide.

📖 Process-GPT User Manual