The problem arrives quietly: an AI agent invokes the incorrect API endpoint, receives a valid response, and proceeds without incident. A monitoring dashboard stays green. For engineering teams trained to find trouble by hunting for failed requests, this kind of success creates confusion.

"In the agentic AI era, we are not querying like humans anymore," said Bianca Lewis, executive director of the OpenSearch Software Foundation.

Lewis made these remarks during a conversation with Frederic Lardinois, senior editor for AI at The New Stack, at AGNTCon + MCPCon North America 2026 in San Jose. She outlined how OpenSearch, the open-source search and analytics platform, fits into both sides of an agent's operation—retrieving the data agents need while giving developers visibility into what agents do with it.

Following the agent's path

Lewis recalls an earlier era when failures announced themselves plainly.

Error 500 queries used to be the best thing that could ever happen because we would know that [something] went wrong. 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. It's changed how we need think about our applications and how we need to observe them.

Bianca Lewis

When an agent has access to multiple services, each API call can complete successfully in isolation, yet the sequence of calls can steer the agent away from what the user actually requested.

Lewis describes pairing OpenSearch with OpenTelemetry to record which services and APIs an agent contacted. By adding cost tracking to the trace, developers gain insight into both where expenses accumulated and where the agent's path diverged from the intended goal. This concrete data becomes essential when the final result appears reasonable but conventional error logs offer no clues.

Lewis also separates the temporary context of a single interaction from the persistent information stored in search indexes. This distinction matters when teams decide what to fix after a failed run. The trace documents what the agent actually executed. Developers can then examine the context the application initially provided to the agent.

Combining search and observability on a single platform appeals most to teams already using OpenSearch for one function. Lewis notes that some teams begin with observability, then apply the same infrastructure to build search capabilities. This approach lets them leverage existing expertise and reduce how many systems store copies of the same data.

When Lardinois raised the topic of data gravity, Lewis countered that the expense of migrating platforms should factor into total cost of ownership calculations alongside operational expenses.

Building the next generation

Platform transitions require time. Teams run production systems for years, maintain existing contracts, and need to build new capabilities. Lewis and Lardinois explored how commercial support and hosting options help teams transition, and how an open-source project gives teams influence over the features they will rely on.

I think that contribution into OpenSearch can take many forms.

Bianca Lewis

Lewis identifies pull requests, conversations in community Slack channels, and user groups as entry points for participation. For junior developers, contributing to OpenSearch builds a visible record of their work on production systems.

Experienced engineers bring systems-level thinking that becomes invaluable when an agent's assured answer signals that something has gone wrong.

Lewis describes this transition with a phrase she uses in jest. "I always joke — and I can say this because of my age — is that it's sort of like revenge of the dinosaurs. We have now come to the stage where the person who understands the system-level context of what you're trying to achieve with agentic AI becomes extremely important."

The full interview is available on The New Stack's YouTube channel, covering what teams are building with OpenSearch, community priorities, and Lewis's perspective on maintaining data control as agent workloads expand.