Overview
Research addresses the challenge of inspecting, debugging, and evaluating Large Language Model (LLM)-based multi-agent systems, particularly under controlled failure conditions, within software engineering contexts. A lightweight, framework-agnostic tool, named llmmas-otel, was developed to integrate OpenTelemetry-based distributed tracing with fault injection capabilities specifically for these systems. The objective of llmmas-otel is to enable reproducible comparison between baseline and faulty executions and to facilitate the inspection of the effects of faults through aligned traces and run artifacts.
Research Context
LLM-based multi-agent systems are increasingly investigated for applications in software engineering tasks. A recognized difficulty associated with these systems is their inspectability, debuggability, and the ability to evaluate their performance and robustness when subjected to controlled failures. The development of methods and tools to address these issues is therefore pertinent to their advancement and practical deployment.
Approach
The core of the research involves the design and implementation of llmmas-otel. This tool is characterized by its use of OpenTelemetry for distributed tracing and its support for fault injection. Key aspects of its approach include:
- **Instrumentation:** llmmas-otel instruments agent executions, gathering trace-aligned telemetry across various operational stages. These stages encompass workflow phases, individual agent steps, inter-agent communication, calls to external tools, and invocations of LLMs.
- **Fault Injection:** The tool supports targeted fault injection. This capability allows for the introduction of faults at selected interaction points within the multi-agent system's operation.
- **Reproducible Comparison:** By combining instrumentation and fault injection, llmmas-otel enables the comparison of system behavior under normal (baseline) conditions versus conditions with induced faults. The inspection of these effects is facilitated through the generated aligned traces and other run artifacts.
The research describes the motivation behind the tool's creation, its architectural design, implementation details, and its current capabilities.
Findings
Initial validation of llmmas-otel was conducted in two distinct environments:
- A minimal demo workflow was used to assess the tool's foundational functionality.
- A real LLM-based multi-agent system designed for software development was employed for further evaluation.
The tool's application in these contexts made it possible to compare baseline and faulty executions in a reproducible manner and to inspect the effects of injected faults using aligned traces and generated run artifacts.
Why This Matters
The ability to inspect, debug, and evaluate LLM-based multi-agent systems under controlled failure conditions is difficult. The introduction of llmmas-otel addresses these challenges by providing a mechanism for combining distributed tracing with fault injection. This capability supports the analysis of how faults impact system behavior across different operational phases, inter-agent communications, tool calls, and LLM interactions, thereby contributing to a better understanding and potentially more robust design of such systems.