Trace Cascading Decision Failures in Multi-Agent LLM Systems with OpenSearch Service
This AWS blog post details a solution for attributing root causes to failures in multi-agent LLM systems, which are typically hard to debug with standard logs. It showcases how to build a blame graph using Amazon OpenSearch Service for graph storage and Amazon Bedrock for embeddings and reasoning. This approach helps engineers and architects identify which agent caused a bad outcome by measuring influence between agents. Prerequisites and source code are provided for implementing the demonstration pipeline.
Notes (1) ›
- Building a Blame Graph for Multi-Agent LLM Debugging
The solution addresses the challenge of debugging multi-agent LLM systems by treating agent reasoning as a graph. It measures influence between agents using embedding similarity and traces backward from a failed decision to identify the root cause. This approach leverages Amazon OpenSearch Service for graph storage and analytics, alongside Amazon Bedrock for text embeddings (Titan Text Embeddings V2) and agent reasoning (Anthropic Claude Sonnet 4.5).
https://aws.amazon.com/blogs/big-data/trace-cascading-decision-failures-with-a-blame-graph-on-amazon-opensearch-service/
Related releases
- Amazon Bedrock AgentCore Memory now extracts long-term memories from JSON payloads AWS What's New ·
- OpenAI GPT-5.6 Terra and Luna Models Now Available in AWS GovCloud AWS What's New ·
- Amazon Bedrock announces reduced pricing for OpenAI GPT-5.6 Sol AWS What's New ·
- Amazon Bedrock AgentCore Web Search Adds Domain/Date Filtering and Regional Expansion AWS What's New ·
- AWS details external web access for Amazon Bedrock's Web Search AWS What's New ·
- Amazon Bedrock now supports SpaceXAI Grok 4.6 with Cross-Region Inference AWS What's New ·