agentuniverse-ai/agentUniverse

agentUniverse is a LLM multi-agent framework that allows developers to easily build multi-agent applications.

agentUniverse – A Multi‑Agent Framework for LLM‑Powered Applications

What it is – agentUniverse is an open‑source Python library (≥ 3.10) that lets you assemble intelligent agents and agent teams built on large language models (LLMs). It supplies a collection of ready‑made collaboration patterns (e.g., PEER – Plan/Execute/Express/Review, DOE – Data‑finding/Opinion‑inject/Express) and a plug‑in style architecture for adding tools, knowledge bases, memory, and observability. The project originates from Ant Group’s real‑world financial‑service use cases and is packaged on PyPI (v0.0.19, Apache‑2.0).

Core Concepts

Concept What it gives you
Agent A configurable LLM‑backed unit with its own prompt, tools, memory, and optional domain knowledge.
Pattern (Collaboration Model) Pre‑defined orchestration of multiple agents that work together on a task (e.g., PEER splits a problem into planning → execution → expression → review).
MCP Server A lightweight service for exposing agents over HTTP, making them callable from other services or UI components.
Observability Built‑in OpenTelemetry hooks that emit traces/metrics for agents, LLM calls, and tool usage.
Visual Workflow Canvas A UI (via the magent-ui package) where you can drag‑and‑drop agents and connect them to form a workflow without writing code.

Key Features (as described in the README)

  • Multi‑Agent Collaboration Patterns – PEER, DOE and a framework for defining your own patterns.
  • Broad LLM Support – Simple configuration lets you switch among dozens of providers (Qwen, DeepSeek, OpenAI, Claude, Gemini, Llama, Kimi, WenXin, ChatGLM, Baichuan, Doubao, etc.).
  • Domain‑Expert Integration – Prompt templates, SOPs, and knowledge‑base hooks let you inject industry expertise (finance, legal, code generation, etc.).
  • Extensible Tooling – Attach arbitrary Python tools, RAG pipelines, or external APIs to agents.
  • Standardised Observability – OpenTelemetry‑compatible tracing for end‑to‑end monitoring.
  • MCP Server – Publish agents as micro‑services; also supports consuming external MCP servers.
  • Visual Canvas – magent-ui provides a no‑code canvas for building and running agentic workflows.
  • Documentation & Samples – Quick‑start guide, step‑by‑step tutorials, and several ready‑made example apps (legal advice, code runner, financial event analysis, etc.).

Typical Use Cases

Domain Example from the repo
Finance / Investment Research Zhi Xiao Zhu AI assistant for analysts (PEER‑driven report generation, ESG analysis).
Legal Legal‑advice agent that parses user queries and returns structured counsel.
Software Development Python code generation & execution agent that writes, runs, and returns results.
Translation / Knowledge Transfer Reflexive workflow translation agent (replicating Andrew Ng’s demo).
Discussion / Brainstorming Multi‑turn, multi‑agent discussion group that simulates a panel of experts.

Getting Started

# Install the library
pip install agentUniverse

# (Optional) Install the visual UI package
pip install magent-ui ruamel.yaml
  1. Configure an LLM – add your provider key to custom_key.toml and set llm_model in the agent’s config (e.g., default_deepseek_llm).
  2. Run the tutorial – follow docs/guidebook/en/Get_Start/2.Run_Your_First_Tutorial_Example.md to launch a sample agent.
  3. Explore samples – the examples/ folder contains a standard‑project scaffold and ready‑made apps.
  4. Launch the visual canvas – run product_application.py under examples/sample_apps/workflow_agent_app/bootstrap/platform.

Documentation & Resources

  • User Guide – docs/guidebook/en/Contents.md (installation, building agents, patterns, observability, etc.)
  • API Reference – hosted at https://agentuniverse.readthedocs.io/en/latest/
  • Research Paper – PEER: Expertizing Domain‑Specific Tasks with a Multi‑Agent Framework and Tuning Methods (arXiv 2407.06985) – the paper that validates the PEER pattern.
  • Community – GitHub Issues, Discord channel, and a public Twitter handle @agentuniverse_ for announcements.

License & Availability

  • License: Apache‑2.0 (per LICENSE file)
  • Package: Published on PyPI (agentUniverse), source code on GitHub under the agentuniverse-ai organization.

Quick Verdict

agentUniverse is a genuine, production‑oriented framework for building LLM‑driven multi‑agent systems, especially aimed at enterprises that need to embed domain expertise (finance, legal, etc.) into AI assistants. Its emphasis on reusable collaboration patterns, extensibility, and observability makes it a solid choice for developers looking to move beyond single‑agent chatbots toward coordinated, task‑oriented AI workflows.

Related

  • Project
    modelscope/ms-agent

    A lightweight framework for building autonomous agents capable of deep research, code generation, and multimodal tasks using a knowledge-driven skill system.

  • Project
    bojieli/ai-agent-book

    AI‑Agent‑Book is a multilingual open‑source textbook on building LLM‑based AI agents, paired with 109 runnable Python experiments covering everything from prompt engineering to multi‑agent collaboration.

  • Project
    dapr/dapr-agents

    Dapr Agents is a Python SDK built on the Dapr runtime that lets developers create scalable, fault‑tolerant AI agents. It provides durable workflows, actor‑based execution, built‑in data source bindings, multi‑agent communication, and automatic tool discovery, all with Kubernetes‑native deployment and strong security.

  • Project
    K-Dense-AI/agentic-data-scientist

    An adaptive multi-agent framework that automates complex data science workflows by separating planning from execution and utilizing a self-correcting, iterative validation loop.

  • Project
    shibing624/agentica

    Agentica is an open‑source, Apache‑2.0‑licensed Python framework that lets you run LLM‑powered agents locally. It provides a unified engine accessible through a terminal CLI, a local web gateway, and native desktop apps. Agents can spawn sub‑agents, delegate whole processes, and exchange peer messages, enabling multi‑agent collaboration. Built‑in tools (file I/O, code execution, web search) run asynchronously, and a self‑evolving *skill* system automatically captures experience for future reuse. The platform is model‑agnostic (OpenAI, DeepSeek, Claude, Ollama, etc.), supports multimodal inputs, safety guardrails, persistent memory, RAG integration, and offers Python and TypeScript SDKs for embedding agents in code. Installation is via `pip install agentica` (CLI) or `pip install "agentica[gateway]"` (web UI), with Docker and pre‑built desktop binaries also available.