AgentVerse

AgentVerse is an open-source framework from OpenBMB for deploying multiple LLM-based agents. It provides two frameworks, task-solving and simulation, and is installable from source or via pip. The repository is released under the Apache 2.0 license.

What It Is

AgentVerse is a multi-agent framework for large language models. Its task-solving framework assembles several agents into an automatic multi-agent system that collaborates on a task, with applications such as software development and consulting systems. Its simulation framework lets users set up custom environments to observe, or interact with, multiple agents, for uses such as games and social behavior research on LLM-based agents.

How It Works

Users install the package, export an OpenAI or Azure OpenAI API key, and run tasks from the command line. Simulation tasks, such as a nine-agent NLP classroom with one professor and eight students, run through the agentverse-simulation command, and a local web GUI is available through agentverse-simulation-gui. Task-solving runs use agentverse-tasksolving for single queries and agentverse-benchmark for benchmark datasets such as HumanEval. Tool-using task-solving cases rely on a ToolServer built from XAgent, and simulation cases with tools optionally use BMTools.

Model Access

Besides OpenAI and Azure OpenAI, the README documents local model support. Local models such as LLaMA and Vicuna can be served through FastChat, and vLLM servers can be connected through environment variables and a model setting in the task config file.

Current Status

The README notes that the code is being refactored, and points to the release-0.1 branch for a stable version that supports only the simulation framework. Its roadmap lists documentation and more sophisticated conversation-history memory as pending items.