Emiyaaaaa/HiveMind

What it solves

Agent frameworks (like LangGraph or AutoGen) are often designed for local prototyping and lack the production-grade infrastructure needed for scalable services. HiveMind (AgentFlow) provides a runtime layer that handles the "operational" side of multi-agent systems—such as durable history, event streaming, and execution monitoring—without forcing the developer to commit to a single orchestration framework.

How it works

It uses a split-runtime architecture to separate API management from agent execution:

• API Tier: A Java/Spring Boot server handles REST requests and bridges Server-Sent Events (SSE) to the frontend.
• Execution Tier: Python workers consume jobs from a Redis queue and execute agents via "Orchestrator Adapters."
• Persistence: All run states, messages, and tool calls are stored as first-class entities in a Postgres database via SQLAlchemy.
• Adapters: A pluggable interface allows different frameworks (e.g., LangGraph, PydanticAI) to be wrapped so they all report status and data in a consistent format.

Who it’s for

Teams moving from agent prototypes to inspectable production services who need a stable abstraction for switching orchestration frameworks and a way to monitor active and historical runs.

Highlights

• Framework Agnostic: Supports multiple orchestrators (LangGraph, PydanticAI, etc.) through a unified adapter interface.
• Durable State: Maintains a persistent record of runs, steps, messages, and checkpoints across process restarts.
• Live Observability: Includes a Next.js admin console and SSE support for real-time streaming of agent execution events.
• Enterprise-Ready Stack: Built with Java 21, Python 3.12, Redis, and Postgres for reliability and scalability.