EvoAgentX
Open-source platform that generates and evolutionarily optimizes multi-agent workflows.
About
EvoAgentX tackles the part of multi-agent systems most frameworks leave manual: designing the workflow itself. You state a goal in natural language, its WorkFlowGenerator produces a multi-agent workflow graph, AgentManager instantiates the agents, and the workflow executes. Then the evolving layer goes further, integrating three optimization algorithms (TextGrad, AFlow, MIPRO) to iteratively refine agent prompts, tool configurations, and workflow topologies.
The architecture is layered cleanly: basic components, agent, workflow, evolving, and evaluation. The accompanying paper reports meaningful gains on HotPotQA, MBPP, MATH, and GAIA benchmarks. The code is MIT-licensed Python and pip-installable.
Caveats worth knowing: it is a research-originated project (the repo moved from the EvoAgentX org to ANative-Lab), and the headline benchmark improvements are from the authors' own paper, not independent reproduction. Best suited to researchers and teams experimenting with self-optimizing agent topologies rather than anyone needing a stable production framework today.
What it can do
- Natural-language to multi-agent workflow generation
- Evolutionary optimization of prompts, tools, and topologies (TextGrad, AFlow, MIPRO)
- Five-layer modular architecture with an evaluation layer
- Workflow visualization and save/load
- MIT-licensed Python package
Setup
pip install evoagentx; configure an LLM (OpenAI-compatible config shown in the README); define a goal and call WorkFlowGenerator to build and run the workflow.
Prerequisites
Python 3.10+; an LLM provider API key; comfort reading research code.
Use cases
Works with
Trust score — 6/10
Scored 6/10 (verified tier, metrics-v1). Evidence: 3,384 GitHub stars; repo moved from EvoAgentX org to ANative-Lab; pushed 2026-08-27; MIT license confirmed via LICENSE file; arXiv paper (2507.03616) corroborates architecture and benchmark claims. Manual review: research-originated project; benchmark gains are author-reported, not independently reproduced; activity slower than peers. License: MIT.
★ 3,384 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · 3,384 GitHub stars; repo moved from EvoAgentX org to ANative-Lab; pushed 2026-08-27
- · MIT license confirmed via LICENSE file; arXiv paper (2507.03616) corroborates architecture and benchmark claims
- · manual review: research-originated project; benchmark gains are author-reported, not independently reproduced; activity slower than peers
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.