The Problem
Building multi-agent LLM systems typically means stitching together orchestration logic, tool integrations, and model providers from scratch. OpenAGI addresses this by providing a framework for creating autonomous, human-like agents with a structured architecture for planning, execution, and tool use.
What This Does
OpenAGI is a Python framework for building multi-agent systems. The core architecture centers on an Admin agent (in src/openagi/agent.py) that coordinates Worker agents, a planner module for task decomposition, and a registry of pre-built actions and tools. The src/openagi/actions/tools/ directory contains 20+ integrations including ddgsearch.py (DuckDuckGo), arxivsearch.py, tavilyqasearch.py, and githubsearchtool.py. LLM providers are abstracted in src/openagi/llms/ with support for OpenAI, Gemini, Claude, Groq, Ollama, and others.
The framework supports both manual multi-agent execution (where an admin delegates to specialized workers) and autonomous execution with a TaskPlanner that decomposes queries without human intervention. A CLI entry point exists at src/openagi/cli.py, and the example/ directory contains 13 runnable use cases from blog writing to job search agents.
How To Use It
Setup: Install via pip install openagi or clone and run pip install -e . from the repo root. Python 3.9–3.11 is supported per the README badge.
Configuration: Set provider API keys as environment variables. The README examples show OPENAIAPIKEY, TAVILYAPIKEY, and GOOGLEAPIKEY. .env.example at the repo root lists available variables.
Running it: Execute any example script directly, e.g., python example/blogpost.py. For custom agents, follow the pattern in the README: instantiate an LLM, create a TaskPlanner, define Worker agents with roles and instructions, assign them to an Admin, then call admin.run(query, description).
pip install openagi or: git clone https://github.com/aiplanethub/openagi.git && pip install -e . python example/blogpost.py
Real-World Use
A practical scenario: a market research agent that searches the web, summarizes findings, and writes a report. The example/marketresearch.py file demonstrates this pattern—combining a search action with an LLM to produce structured output. The cookbook/ directory contains Jupyter notebooks showing custom tool integration and human-in-the-loop intervention, useful for teams that need interactive agent workflows.
Code Health & Issues
Med - Minimal test coverage: Only 1 test file exists for 65 Python files. Core orchestration logic in agent.py and planner/ is effectively untested. Low - Duplicated documentation assets: .gitbook/assets/ and docs/.gitbook/assets/ contain near-identical image sets (40+ images each), bloating the repo. Low - Mixed dependency management: Both requirements.txt and pyproject.toml/poetry.lock are present, which can cause version drift between install methods. Low - CI scope: .github/workflows/prreviewcheck.yml exists but appears limited to PR review checks; no visible test or build pipeline. Med - Error handling: Tool integrations (e.g., googlesearchtool.py, serpsearch.py) likely need API keys but error messages for missing configuration aren't evident from the structure.
The Bottom Line
OpenAGI is a functional multi-agent framework with solid provider coverage and useful example applications. It's a reasonable choice for teams wanting to prototype agent workflows quickly without building orchestration from scratch. The thin test coverage and dual dependency files suggest it's early-stage; treat it as a development framework rather than production infrastructure.