AgentLite: A Lightweight Library for Building and Advancing Task-Oriented LLM Agent System

• Existing libraries can be difficult to refactor because they impose complex agent creation overhead, hard-coded reasoning types, or fixed architectures.LangChain is described as over-complicated for new reasoning types and architectures, while AutoGen fixes multi-agent conversation and code execution.Source: §1 · p. 2
• AgentLite provides a lightweight code architecture with minimal dependencies and customizable prompts, memory, action sequences, and LLMs.Source: §1 · p. 2
• Its task-oriented design assigns agents specific tasks and supports communication and collaboration in complex multi-agent systems.Source: §1 · p. 2
• AgentLite supports hierarchical multi-agent orchestration by using manager agents to coordinate teams of individual agents.The manager-agent arrangement is described as supporting broader task objectives through coordinated team-agent actions.Source: §1 · pp. 2–3
• The library is intended for both benchmark evaluation and deployment in real-world applications, demonstrating practical implementation potential.Source: §1 · p. 3
• AgentLite achieves comprehensive LLM-agent capabilities with less than 1k lines of code, compared with existing libraries including AutoGen, LangChain, Camel, and CrewAI.Source: §1 · p. 3

2 AgentLite Framework

• The framework combines a base Individual Agent with a Manager Agent for constructing multi-agent systems.Source: §2 · p. 3, Figure 1 · p. 3

2.1 Individual Agent

• An Individual Agent uses PromptGen, LLM, Actions, and Memory to generate executable actions, receive observations, and retain action-observation history.Source: Figure 1 · p. 3, §2.1 · pp. 3–4
• PromptGen assembles role, instruction, constraint, action, and example components, while developers can customize prompts for action generation.Source: §2.1 · p. 3
• Developers extend BaseAction with executable functions and parameter documentation; adding actions such as Think enables reasoning strategies such as ReAct.Source: §2.1 · pp. 3–4

2.2 Manager Agent

• ManagerAgent decomposes a task into sub-tasks and delegates them to individual agents, which can themselves be organized into deeper hierarchies.Source: §2.2 · pp. 4–5
• TaskPackages communicate instructions and results between managers and subordinate agents, with later packages generated sequentially from earlier execution feedback.Source: §2.2 · p. 5

3.1 New Reasoning Type

• AgentLite unifies reasoning and executable tool use as Action modules, making new reasoning types such as ReAct easy to add.Developers can also create analogous Plan and Reflect actions by changing action prompts and parameters.Source: §3.1 · pp. 5–6

3.2 New Agent Architecture

• New agent architectures can be configured in a few lines of Python, including systems with human agents and multiple LLM backbones.AgentLite supports arbitrary numbers of agents, human-in-the-loop orchestration, and distinct LLM configurations for different agents.Source: §3.2 · pp. 6–7
• A Copilot Agent receives human instructions through a HumanInput action added to a BaseAgent alongside other actions.The HumanInput action inherits from BaseAction and returns the instruction collected from a human.Source: §3.2 · p. 6
• Copilot Multi-Agent systems add a HumanAgent to a manager-controlled team so human instructions guide other agents' task completion.The HumanAgent subclasses ABCAgent and collects instructions through interactive input.Source: §3.2 · pp. 6–7
• Multi-LLM Multi-Agent systems assign configurable LLM backbones to different agents and assemble them under a manager agent.The manager's own LLM is also configurable through LLMConfig.Source: §3.2 · p. 7

4 Benchmark

• The benchmark section evaluates AgentLite implementations across different agent benchmarks to assess performance with multiple language models.The reported applications include HotPotQA retrieval-augmented question answering and WebShop browser decision making.Source: §4 · p. 7, §4.1 · p. 8, §4.2 · p. 8

4.1 Retrieval-Augmented Question Answering

• HotPotQA agents use WikipediaSearch, Think, and Finish actions to retrieve information, integrate it through reasoning, and produce a final answer.The benchmark evaluates multi-hop question answering across multiple documents.Source: §4.1 · pp. 7–8
• GPT-4 variants outperform GPT-3.5 across all HotPotQA difficulty levels, while GPT-4-32k performs best on the medium set and xLAM-v0.1 exceeds GPT-3.5.The HotPotQA evaluation uses F1-Score and Accuracy across easy, medium, and hard questions.Source: §4.1 · pp. 7–8

4.2 Web Browser Decision Making

• On WebShop, GPT-4-32k performs much better than other LLMs, especially on hard shopping tasks, while xLAM is comparable with GPT-3.5.The evaluation follows AgentBoard's setup with 251 instances: 182 easy and 69 hard tasks.Source: §4.2 · p. 8, Table 3 · p. 8
• WebShop agents require wrapping the environment's actions with BaseAction and inserting them into BaseAgent, then receive final shopping rewards.WebShop contains 1.18M real-world products and associates each instruction with one ground-truth product and attribute requirements.Source: §4.2 · p. 8

5 Applications

• AgentLite applications span practical, human-in-the-loop, multimodal, and innovative settings, demonstrating varied uses of the framework.Source: §5 · p. 9

5.1 A1: Online Painter

• Online Painter coordinates manager, search, and painter agents to retrieve visual information and generate an image from object features.Search uses DuckDuckGo and Wikipedia, while the Painter Agent uses a DALLE action.Source: §5.1 · pp. 9–10

5.2 A2: Interactive Image Understanding

• Interactive Image Understanding combines image display, image querying, and human input for multi-round visual question answering.The ImageQuery action is backed by the gpt-4-vision-preview API.Source: §5.2 · p. 10

5.3 A3: Math Problem Solving

• Math Problem Solving uses a single agent with HumanInput and WolframAlphaSolver actions, showing that simple applications need not use a manager agent.The paper contrasts this with more complex multi-agent scenarios where a Manager Agent organizes the workflow.Source: §5.3 · p. 10

5.4 A4: Chess Game

• Chess Game pairs human keyboard input with a Chess Agent that views the board, finds legal moves, and moves pieces, while also supporting agent-agent play.The Chess Agent uses BoardView, LegalMoves, and BoardMove actions in addition to Think and Finish.Source: §5.4 · pp. 10–11

5.5 A5: Philosophers Chatting

• Philosophers Chatting uses a Manager Agent to ask Socrates, Confucius, and Aristotle agents for opinions and summarize their responses.The application combines a human questioner with philosopher agents grounded by their literary works.Source: §5.5 · p. 11

• AgentLite supports benchmark reproduction, application prototyping, and integration and evaluation of new reasoning strategies and agent architectures.The paper also identifies additional inter-agent communication methods and reasoning types as future work.Source: §6 · p. 11