article --summary
The summary
Traditional monitoring relies on failed requests such as HTTP 500 errors to reveal problems, but AI agents can make mistakes that register as successful calls. Lewis argues that teams need to observe the sequence of an agent's actions instead of individual outcomes. In her view, this is where the observability work matters most.
Lewis describes pairing OpenSearch with OpenTelemetry to trace which services and APIs an agent calls, and adding cost data so developers can see where a run got expensive and where it went off course. The trace can then guide changes to the context the application gives the agent, separating short-lived interaction context from longer-lived information stored in search indexes. She also notes that teams already using OpenSearch for observability can extend it to search, reducing duplicate copies of business data.
The conversation also covers the cost of switching platforms, which Lewis counts as part of total cost of ownership, and how community contributions through pull requests, Slack discussions, and user groups can shape the project's direction. She stresses that engineers who understand system-level goals remain essential as agentic AI spreads.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.