How TReNDS automates root-cause analysis with Amazon Bedrock
AWS Vitaly Omelchenko
A Georgia State research team hooked Claude into their AWS logging pipeline to auto-diagnose production errors. It turns 15-30 minute debugging sessions into sub-60-second reports with a suggested fix already attached.
Based on reporting by AWS, Vitaly Omelchenko — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Root-cause analysis is one of those engineering tasks everyone hates and nobody automates well, because it requires actually understanding code, not just matching error strings. The team at TReNDS, the neuroimaging research center jointly run by Georgia State, Georgia Tech, and Emory, decided to try anyway, and they wrote up how it went in a recent AWS blog post co-authored with engineer Vitaly Omelchenko.
Their setup is refreshingly unglamorous. Applications running on Amazon EKS ship logs to CloudWatch through FluentBit, same as thousands of other AWS shops. The twist is a CloudWatch subscription filter watching for ERROR, Exception, FATAL, and CRITICAL patterns, which fires a Lambda function the moment something breaks. That Lambda runs an agent built with the Strands Agents SDK, powered by Claude Sonnet 4 through Amazon Bedrock, and the agent does what a tired on-call engineer would do at 2am: pull surrounding log context, fetch the actual source file from GitHub, trace the stack, and figure out what went wrong.
The engineering detail worth noting is that the team didn't hardcode an investigation flowchart. They gave the model tools — a fetch_source_code function that hits the GitHub API, a fetch_log_context function that scopes CloudWatch queries to the exact container stream that threw the error — and let Claude decide which to call and when. A clean stack trace triggers a source-code fetch. A vaguer error might send the agent searching the codebase for a matching string instead. That's the whole pitch of agentic tooling: you stop trying to anticipate every failure mode and let the model improvise within guardrails set by a system prompt asking for severity, root cause, code context, a suggested fix, and related risk areas.
Because TReNDS handles health research data that can fall under HIPAA, keeping everything inside their own AWS account mattered as much as the automation itself. Bedrock processes the requests without shipping logs or source code to some third-party endpoint, which let the team build this without opening a new compliance headache.
The numbers are the real story. Investigation time went from 15-30 minutes, sometimes much longer for cross-service issues, down to under a minute. Engineers now often get a diagnosis and a proposed fix sitting in their inbox before they've finished their coffee. Deduplication through DynamoDB keeps repeat errors from spamming the channel, and the whole thing runs cheap — two to three tool-use rounds per error, which the team calls a rounding error next to the cost of engineer time.
My take — AI-written commentary, not fact-checked reporting
This is exactly the boring, unsexy use of agents that actually pays off, and it's a useful counterweight to the endless demos of AI writing entire apps from scratch. Nobody needs a model that can build a startup in a weekend; plenty of teams need one that reads a stack trace, checks the actual file on line 142, and tells them what broke. The HIPAA-conscious, everything-stays-in-your-own-account design is also the template more research and healthcare shops should be copying instead of chasing whatever chatbot is trending that week.
Read more about this at: AWS