We were back at Carahsoft in Reston on October 22 for the third Washington DC OpenSearch User Group meetup. We had about 25 attendees. The room was a good mix of people with various levels of OpenSearch experience. @rutger and I started with a few announcements and then handed it over.
Leonardo (Leo) Segovia, Sr. Search Solutions Architect on Amazon OpenSearch Service, presented “Teaching your Agent to Read a Cluster: Agentic performance analysis for OpenSearch, from MCP to source graphs to a knowledge base of the whole project.” Leo works with strategic customers running OpenSearch at volume, where he spends a lot of his time chasing race conditions that only show up at production scale. He came to OpenSearch the long way, after years of building observability by hand on Nagios, Graylog and the ELK stack.
He started from a simple observation. Most engineers who use OpenSearch don’t operate it. When performance degrades, the evidence is already in the APIs, metrics and logs, but those all assume you know which one to read. So Leo built an agentic analysis workflow live, one layer at a time, using pieces that already ship with the project. First, the official OpenSearch MCP server and the agent skills repo. You install one package, point it at a cluster, add a few lines of agent config and start asking questions in plain English. There’s no plugin to build, no model to tune and nothing to change on the cluster. Second, knowledge graphs built over the OpenSearch source tree and its Apache Lucene dependency, so the agent can explain why a symptom happens instead of only reporting it. Third, a Bedrock knowledge base backed by an OpenSearch vector store, holding the full OpenSearch documentation and Lucene query cost semantics.
For me, the highlight of the night was watching those layers go on one at a time. With each one, the agent had more to work with: first the cluster’s own signals, then the code paths behind a symptom, then answers grounded in the docs and Lucene’s cost model. Leo closed with how to package all of it into a repeatable workflow. The goal is accurate root cause analysis without having to become an OpenSearch internals expert first.
Thanks to Leo for a great talk and live demo, to @kris and the OpenSearch Project for the food, to Carahsoft for hosting us again, to Qmulos and OpenSearch for the swag, and to @rutger for co-organizing. Thank you to OpenSearch Ambassador Eric Pugh for quickly introducing the new OpenSearch Contribution Ladder structure. And thanks to everyone who came out and stayed for Q&A.
What is next
We are targeting December for meetup number four. Details are coming on Meetup, so join the group to get the announcement: OpenSearch Project Washington DC | Meetup
We are also looking for speakers. If you run OpenSearch in production or have built something on top of it, we’d like to hear from you. Migration, relevance and performance stories are all good topics, and first-time speakers and lightning talks are welcome. Reply here or message Rutger or me on Meetup.
See you in December!