AgentFlow provides a Python-first runtime layer for multi-agent systems, decoupling orchestration frameworks from essential production concerns like persistent run state, event streaming, and operational tooling. It targets teams transitioning from agent prototypes to inspectable, scalable services, offering a consistent execution model and enabling flexibility across orchestration frameworks like LangGraph, AutoGen, or CrewAI.

How It Works

The system employs a dual-backend architecture for scalability and separation of concerns. A Java/Spring Boot API serves the Next.js frontend, handling REST requests and Server-Sent Events (SSE) for real-time updates. Agent orchestration is managed by Python worker processes that consume jobs from a Redis queue, driving pluggable adapters (e.g., LangGraph) to execute agent logic. All state, including runs, steps, messages, and tool calls, is persisted in PostgreSQL via SQLAlchemy, providing a unified observability surface across different orchestration engines.

Quick Start & Requirements

Requirements include Docker, uv, and Node.js 20+. The target Java API architecture additionally requires JDK 21 and Maven 3.9+. Two setup options are detailed:

1. Legacy FastAPI: Single-process, in-memory mode using docker compose up -d postgres redis, followed by Python environment setup (uv sync, alembic upgrade head) and frontend setup (npm install, npm run dev).
2. Java API + Python Worker (Target): Requires infrastructure setup (docker compose --profile java up --build), database migrations (alembic upgrade head), running the worker in queue mode (AGENTFLOW_WORKER_MODE=queue), and separate shells for the Java API (mvn spring-boot:run) and frontend (npm run dev). Example curl commands are provided for creating agents and initiating runs.

Highlighted Details

• Pluggable Orchestrator Adapters: Supports integrating various agent frameworks (LangGraph, AutoGen, CrewAI, PydanticAI) via a consistent adapter interface without altering the core API or database schema.
• Persistent Execution Model: Run, Step, Message, ToolCall, and Checkpoint are first-class database entities, ensuring durable run history and a common observability surface.
• Server-Sent Events (SSE): Emits real-time lifecycle events, allowing clients to stream execution progress without polling.
• Lightweight Admin Console: A Next.js-based console provides run listing, detail views, step rendering, and live event stream subscription.

Maintenance & Community

The README outlines a roadmap with planned phases for enhanced observability, extensibility (plugins, SDKs), and production features (workflows, RBAC, OpenTelemetry). Specific details on active contributors, sponsorships, or dedicated community channels (e.g., Discord, Slack) are not provided. Contributions are encouraged via issue reporting, adapter implementation, and console improvements.

Licensing & Compatibility

The project is licensed under the Apache 2.0 license, which is permissive and generally suitable for commercial use and integration into closed-source applications.

Limitations & Caveats

The target production architecture involves both Java and Python components. While Phase 1 (Runtime core) is complete, advanced features such as step timelines, retry/resume capabilities, token/cost summaries, plugin registries, and Temporal-backed workflows are slated for future development phases.