Source-linked AI summary
Agents: An Open-source Framework for Autonomous Language Agents
Wangchunshu Zhou, Yuchen Eleanor Jiang, Long Li, Jialong Wu, Tiannan Wang, Shi Qiu, Jintian Zhang, Jing Chen, Ruipu Wu, Shuai Wang, Shiding Zhu, Jiyu Chen, Wentao Zhang, Xiangru Tang, Ningyu Zhang, Huajun Chen, Peng Cui, Mrinmaya Sachan
TL;DR
Existing language-agent demos are difficult to customize and tune, and often cover only a subset of core capabilities. AGENTS is an open-source framework that combines symbolic control with support for tools, memory, and multi-agent interaction. Case studies demonstrate single-agent, multi-agent, and human-agent systems built with the library.
Problem
Many language-agent demos are hard to customize and tune, cover only a subset of core abilities, and rely on LLM planning that can yield run-to-run randomness and inconsistency.
Source: §1 · pp. 1–2Method
AGENTS provides an open-source framework with symbolic plans, tool and web APIs, human-agent interaction, and an automated retrieval-augmented pipeline for generating configs from short task descriptions.
Source: §1 · pp. 2–3Results
Case studies demonstrate AGENTS-built single-agent, multi-agent, and human-agent interaction systems.
Source: §4 · p. 8Takeaways & Limitations
Abstract
from arXiv · showhide
Recent advances on large language models (LLMs) enable researchers and developers to build autonomous language agents that can automatically solve various tasks and interact with environments, humans, and other agents using natural language interfaces. We consider language agents as a promising direction towards artificial general intelligence and release Agents, an open-source library with the goal of opening up these advances to a wider non-specialist audience. Agents is carefully engineered to support important features including planning, memory, tool usage, multi-agent communication, and fine-grained symbolic control. Agents is user-friendly as it enables non-specialists to build, customize, test, tune, and deploy state-of-the-art autonomous language agents without much coding. The library is also research-friendly as its modularized design makes it easily extensible for researchers. Agents is available at https://github.com/aiwaves-cn/agents.
1 Introduction
Existing language-agent demos are difficult to customize and tune, and many cover only a subset of core agent abilities. AGENTS is an open-source framework designed to make building and customizing agents accessible while supporting finer control and multi-agent use.
1 Introduction
- Existing demos are difficult to customize and tune, cover only a small portion of core agent abilities, and can produce run-to-run randomness and inconsistency.The paper identifies these limitations as barriers to broader access and a better user experience.Source: §1 · pp. 1–2
- AGENTS is an open-source framework intended to simplify customizing, tuning, and deploying language agents for non-specialists while remaining extensible for researchers and developers.The project also provides documentation and tutorials on GitHub.Source: §1 · pp. 2–3
- AGENTS supports external tools through common APIs and an extensible abstract class, and offers specialized APIs for web search and navigation.It also supports long- and short-term memory, which users can select through a config field; the memory components are integrated from.Source: §1 · p. 2
- Dynamic scheduling uses a moderator controller to choose the next agent based on agent roles and interaction history, with rules configurable in natural language.The framework also supports customizing multi-agent systems for games, social experiments, and software development.Source: §1 · pp. 2–3
- AGENTS supports human communication and interaction with language agents in both single-agent and multi-agent scenarios.The paper identifies the difficulty of human interaction, especially in multi-agent settings, as a limitation of existing frameworks.Source: §1 · p. 3
- Symbolic plans, or SOPs, represent task situations as states with transition rules and step-by-step instructions, enabling fine-grained control and potentially more stable, predictable behavior.An SOP can be generated by an LLM and edited by users; agents adjust their current state in response to the environment, humans, or other agents.Source: §1 · p. 3
- An automated SOP-generation pipeline uses retrieval-augmented generation to create agent config files from a short task description, reducing the labor of writing detailed SOPs and configs.The pipeline is described as a “meta agent.”Source: §1 · p. 3
2 Related Work
Related work surveys language agents built for specific tasks and open-source frameworks for broader agent construction. It contrasts framework capabilities with AGENTS’ support for tool use, memory, multi-agent communication, human interaction, and SOP-based controllability.
2.1 Autonomous Language Agents
- Generative Agents, WebAgent, Qian et al. [2023], MetaGPT, and Zhou et al. [2023a] target social behavior, websites, software development, and writing assistance.These examples represent language agents designed for different tasks.Source: §2.1 · p. 3
- AutoGPT [Richards and et al., 2023], BabyAGI, and SuperAGI pursue agents that can handle varied user requests and drew interest from developers and non-specialists.Source: §2.1 · p. 3
2.2 Language Agents Frameworks
- Frameworks include tool-using Transformers Agents, end-to-end LangChain, multi-agent platforms Camel and AgentVerse, and tool-augmented libraries Gentopia and XLang4.Source: §2.2 · pp. 3–4
- AGENTS is described as the only compared framework supporting tool usage, long-short term memory, and multi-agent communication together, while also offering human-agent interaction and SOP-based controllability.The comparison is presented in Table 1.Source: §2.2 · p. 4, Table 1 · p. 4
3 Library Design
AGENTS organizes agents, symbolic plans, and environments around a configurable, extensible framework, then adds tools, memory, multi-agent coordination, human interaction, deployment, sharing, and automatic SOP generation.
3 Library Design
- AGENTS initializes agents, SOPs, and environments from a plain-text config, with modular prompts and standardized APIs that developers can override.The config also supports sharing language-agent systems.Source: §3 · p. 4
3.1 Agent
- The Agent class observes the environment, acts, and updates memory; an is_human setting lets a human take an agent’s role.Developers can customize the factorized agent methods with new functionality.Source: §3.1 · p. 5
3.2 SOP
- An SOP models agent states and subtasks as a graph, with modular prompt and tool components for each state and LLM-based state transitions and agent routing.PromptComponents specify goals, constraints, demonstrations, and output format; ToolComponents support APIs and external tools.Source: §3.2 · p. 5
3.3 Environment
- The Environment class defines what agents observe and how their actions update the environment, while the execution loop can add states for dynamic replanning.Each iteration selects a state and agent, runs an action, and updates the environment.Source: §3.3 · pp. 5–6
3.4 Implementation Details of Core Features
- AGENTS implements long-term memory with embedded action histories and semantic search, supports tools and web search through ToolComponents, and uses an LLM controller for flexible agent scheduling.Short-term working memory is updated by an LLM; human users can also act as agents by setting is_human to True.Source: §3.4 · p. 6
3.5 Deployment
- AGENTS supports deploying language agents as FastAPI APIs, facilitating integration into real-world applications.Existing open-source frameworks were described as focusing on terminal or Gradio proof-of-concept agents.Source: §3.5 · p. 7
3.6 The Agent Hub
- Agent Hub lets users share, search, and download fine-tuned agents, which they can adapt instead of building and tuning from scratch.Users can start from community agents and make modifications.Source: §3.6 · p. 7
3.7 Automatic Creation of Agent Systems
- Automatic SOP generation uses retrieval-augmented generation in a staged pipeline that specifies agents, plans states and connections, then generates components.The pipeline itself follows an SOP and is described as a “meta agent” for creating agents and multi-agent systems.Source: §3.7 · p. 7
4 Case Studies
The case studies demonstrate AGENTS in single-agent, multi-agent, and human-agent settings, including varied tasks and cooperative and competitive scenarios.
4 Case Studies
- The case studies span single-agent systems, multi-agent systems, and systems requiring human-agent interaction.The demos are available on the project website.Source: §4 · p. 8
4.1 Single-agent systems
- Single-agent examples include a chit-chat bot, two knowledge-base and web-search customer-service agents, a shopping assistant, and a sales agent.The examples illustrate different library features and use cases.Source: §4.1 · p. 8
4.2 Multi-agent systems
- Multi-agent case studies cover a fiction studio, debate, and software company, with cooperative and competitive scenarios and SOP-controlled subtasks.The examples show how prompts can be inspected and engineered to customize and improve the system; one debate case includes a human participant.Source: §4.2 · p. 8
5 Conclusion
AGENTS is presented as a unified open-source framework and library for language agents, intended to serve developers, researchers, and non-technical audiences.
5 Conclusion
- AGENTS is a unified framework and open-source library for language agents.Source: §5 · p. 8
- The framework aims to help developers build applications with language agents and researchers conduct language-agent research.Source: §5 · p. 8
- AGENTS aims to let general non-technical audiences build and customize personalized language agents.Source: §5 · p. 8