Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ReDel: A Toolkit for LLM-Powered Recursive Multi-Agent Systems

Andrew Zhu, Liam Dugan, Chris Callison-Burch

TL;DR

ReDel is introduced: a toolkit for recursive multi-agent systems that supports custom tool-use, delegation schemes, event-based logging, and interactive replay in an easy-to-use web interface and is able to achieve significant performance gains on agentic benchmarks and easily identify potential areas of improvements through the visualization and debugging tools.

Abstract

Recently, there has been increasing interest in using Large Language Models (LLMs) to construct complex multi-agent systems to perform tasks such as compiling literature reviews, drafting consumer reports, and planning vacations. Many tools and libraries exist for helping create such systems, however none support recursive multi-agent systems -- where the models themselves flexibly decide when to delegate tasks and how to organize their delegation structure. In this work, we introduce ReDel: a toolkit for recursive multi-agent systems that supports custom tool-use, delegation schemes, event-based logging, and interactive replay in an easy-to-use web interface. We show that, using ReDel, we are able to easily identify potential areas of improvements through the visualization and debugging tools. Our code, documentation, and PyPI package are open-source and free to use under the MIT license at https://github.com/zhudotexe/redel.

ReDel: A Toolkit for LLM-Powered Recursive Multi-Agent Systems

TL;DR

ReDel is introduced: a toolkit for recursive multi-agent systems that supports custom tool-use, delegation schemes, event-based logging, and interactive replay in an easy-to-use web interface and is able to achieve significant performance gains on agentic benchmarks and easily identify potential areas of improvements through the visualization and debugging tools.

Abstract

Recently, there has been increasing interest in using Large Language Models (LLMs) to construct complex multi-agent systems to perform tasks such as compiling literature reviews, drafting consumer reports, and planning vacations. Many tools and libraries exist for helping create such systems, however none support recursive multi-agent systems -- where the models themselves flexibly decide when to delegate tasks and how to organize their delegation structure. In this work, we introduce ReDel: a toolkit for recursive multi-agent systems that supports custom tool-use, delegation schemes, event-based logging, and interactive replay in an easy-to-use web interface. We show that, using ReDel, we are able to easily identify potential areas of improvements through the visualization and debugging tools. Our code, documentation, and PyPI package are open-source and free to use under the MIT license at https://github.com/zhudotexe/redel.
Paper Structure (26 sections, 7 figures, 6 tables)

This paper contains 26 sections, 7 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Recursive Multi-Agent Systems.
  4. Multi-Agent System Frameworks.
  5. System Design
  6. Tool Usage
  7. Delegation Schemes
  8. Events & Logging
  9. Web Interface
  10. Home Page.
  11. Interactive View.
  12. Save Browser.
  13. Replay View.
  14. Evaluation & Case Study
  15. Experimental Setup
  16. ...and 11 more sections

Figures (7)

  • Figure 1: ReDel allows developers to create systems of recursive agents, inspect each agent's state, and visualize a system's delegation graph (right). Recursive agents can be used to solve complex tasks, such as planning a trip to Japan (left).
  • Figure 2: An example of a simple ReDel tool that exposes an HTTP GET function to any agent equipped with the tool.
  • Figure 3: Using ReDel to define a custom event and dispatch it from a tool. Custom events can be used to add observability deep within a system and can be queried post-hoc for rich data analysis.
  • Figure 4: Every event in a ReDel system, builtin or custom, is logged to a JSONL file. Developers can use data analysis tools of their choice to analyze event logs post-hoc. This example demonstrates token counting.
  • Figure 5: The four views of the ReDel web interface: Home (a), Interactive (b), Save Browser (c), and Replay (d).
  • ...and 2 more figures