Asked what the OpenSearch community is building for agentic workloads, Lewis argues agent mistakes fail silently, which makes agent traces a new kind of telemetry that observability tools need to handle.
…I give you confident answers. Now, in the old days, when we used to control our observability on our search applications, we used to see something failing or an error message. You know, error 500 queries used to be the best thing that could ever happen because we would know that that went wrong.
It's true. It's a different approach that I've taken to seeing those, but yes.
And I like an agent is not going to give me back an error 500 message anymore because it might erroneously call the wrong API, but it's not going to be an error. It's not going to show up as an error. So it's changed how we need to think about our applications and how we need to observe them. So Agentic AI is generating agents which become a new telemetry. So OpenSearch is naturally integrated with OpenTelemetry. And part of that is actually monitoring the AI agents themselves. And we should enrich that with cost data. We should see the full trace history of these AI agents, what services and what APIs they are calling, what they shouldn't be calling. And that's how we can use our observability to inform the AI agents and the context in the actual search applications.
Yeah, context is the key there, right?
Correct. Correct. Because, you know, there's a transactional short-term context and there's a longer-term context which should live in the indexes.
Yeah.