EvoAgentX is an open-source framework designed to build, evaluate, and evolve LLM-based agents or agent workflows in an automated, modular, and goal-driven way. It enables developers and researchers to go beyond static prompt chains or manual workflow orchestration by introducing a self-evolving agent ecosystem that constructs, evaluates, and optimizes AI agents through iterative feedback loops, similar to the process of continuous software testing and improvement.
Core capabilities include:
- Automatic construction of agent workflows: Starting from a single prompt, EvoAgentX can build a structured multi-agent workflow tailored to the task.
- Built-in evaluation: Integrated automatic evaluators score agent behavior according to task-specific criteria.
- Self-evolution engine: Agents do not just work; they can also learn. EvoAgentX uses self-evolution algorithms to improve workflows.
- Plug-and-play compatibility: It can easily integrate original OpenAI and Qwen or other popular models, including Claude, Deepseek, kimi, etc., through LiteLLM, siliconflow, or openrouter. If you want to deploy an LLM on your local machine, you can try LiteLLM.
- Comprehensive built-in tools: It provides a rich set of built-in tools that let agents interact with the real environment.
- Memory modules: It supports temporary (short-term) and persistent (long-term) memory systems.
- Human-in-the-loop (HITL) interaction: It supports interactive workflows in which humans review, correct, and guide agent behavior.
EvoAgentX has a wide range of uses and is suitable for AI researchers, workflow engineers, or startup teams to quickly build fully functional agent systems from vague ideas, minimizing engineering effort and maximizing flexibility. Specific scenarios include: automatically evolving and optimizing agent workflows; supervising and controlling processes through HITL checkpoints; using short-term and long-term memory modules to let agents remember, reflect, and improve; automatically assembling multi-agent workflows after describing goals; and enabling agents to interact with the real world through built-in tools such as search, code, browser, file I/O, and APIs.
Installation and usage: pip installation is recommended, and installation from source is also possible. For local development or detailed setup (such as using conda), refer to the Installation Guide. LLM configuration requires setting an API key, which can be set in the Windows Command Prompt or PowerShell, or by creating a .env file in the project root directory and loading it. After configuration is complete, you can initialize the LLM, define natural language goals, and use WorkFlowGenerator Generate a workflow by instantiating agents through AgentManager, then execute it through WorkFlow. Workflows can be visualized (workflow_graph.display()) and saved/loaded (save_module() / from_file()). See workflow_demo.py for a complete example.
Built-in tools cover multiple categories: code interpreters (Python, Docker), search and HTTP requests (Google, Wikipedia, arXiv, RSS), file system tools (read/write, shell commands), databases (MongoDB, PostgreSQL, FAISS), image tools (analysis, generation), and browser automation (low-level and LLM-driven). EvoAgentX also supports MCP tools and provides a configuration tutorial.
Tool-enabled workflow generation: In more advanced scenarios, a toolkit list can be provided to WorkFlowGenerator, and the generator will consider and incorporate agents as appropriate. For example, enable the Arxiv toolkit.
Human-in-the-loop support: EvoAgentX supports integrating human-in-the-loop interaction into agent workflows, and can pause execution for human approval or injection of user input. All human interactions are managed through a central HITLManager instance, including HITLInterceptorAgent (approval gating) and HITLUserInputCollectorAgent (collecting user data). For example, require human approval before the email-sending agent performs an action; when triggered, it will pause and prompt on the console to [a]pprove or [r]eject.
Application demos include: a financial information agent workflow (collect public company information based on a given index, including market indices, stock prices, institutional buying and selling activity, etc., generate an HTML report, and provide buy/sell/hold recommendations); an ArXiv research summary workflow (based on ArXiv MCP tools, retrieve and summarize papers according to keywords and time range).
Evolution algorithms: EvoAgentX integrates several effective agent/workflow evolution algorithms and evaluates optimization performance on multi-hop question answering (HotPotQA), code generation (MBPP), and reasoning (MATH) tasks. It is also used to optimize the performance of existing open-source multi-agent systems on the GAIA benchmark, such as Open Deep Research and OWL.
Project updates include the August 2025 release ofA Survey of Evolving AI Agents, the framework paper went live on arXiv in July 2025, GitHub stars exceeded 1000, and it was officially released in May 2025.
Community support: You can join the community via Discord, X (formerly Twitter), and WeChat. Weekly meetings (Sunday 16:30–17:30 GMT+8) can be added to the calendar. Contact email: evoagentx.ai@gmail.com, with replies within 2-3 business days. The project welcomes contributions and provides contribution guidelines. If you find EvoAgentX helpful, you can cite its paper. The project is built on top of multiple open-source projects, including AFlow, TextGrad, DSPy, EvoPrompt, LiveCodeBench, etc.
Limitations: The materials do not provide specific information such as charges, pricing, platform restrictions, or performance guarantees, and none is currently provided.
