AWS has updated Amazon OpenSearch Serverless to make it easier to use as a search and vector retrieval backend for AI agents. The new version is designed for bursty AI workloads, scaling up during retrieval-heavy tasks and back down when applications are idle.
The core change is elasticity. According to AWS, the next generation of Amazon OpenSearch Serverless can scale from zero to thousands of requests per second and then return to zero when idle. The company said that design is meant to support applications with uneven demand, including AI agents that may generate bursts of retrieval activity while reasoning through a task. AWS said the service can provide up to 60% cost savings compared with OpenSearch Service clusters provisioned for peak capacity.
The update also focuses on faster deployment and capacity scaling. AWS said the next generation of OpenSearch Serverless can create resources in seconds and scale capacity up to 20 times faster than the previous generation. At launch, the new version supports full-text search and vector search collection types. Customers can create collections through the OpenSearch Service console, AWS CLI or AWS SDKs, while AWS is also offering an option to switch back to the existing OpenSearch Serverless infrastructure.
Search used to be about returning the right result for a query. For OpenSearch, agentic AI makes that job messier: the platform is being asked to help agents gather context, remember what matters and decide what to retrieve next. Bobby Mohammed, principal product manager for search, GenAI and agentic AI at AWS, described that changing role during a recent Open Source Summit North America session in Minneapolis. He said OpenSearch has moved from keyword retrieval into semantic and vector search, and is now being adapted again for agentic systems.
Mohammed described how traditional search and observability architectures were largely built around single-turn retrieval. Agentic systems, especially multi-agent systems, create a different set of requirements. Instead of asking one question and returning one set of results, they may reason through multiple steps, call tools, maintain context and revise their next retrieval based on what they find.
That change from single-turn retrieval to agent workflows creates a practical problem for enterprise AI applications: their usefulness depends less on model behavior alone than on the quality of the information the model can reach. Mohammed described the agentic shift as requiring new primitives around retrieval, context, orchestration, monitoring and governance. In that view, OpenSearch is not just returning results but helping supply the context agents need to produce more accurate answers.
AWS is also bringing those capabilities closer to developer workflows. The company said developers can now create a new OpenSearch collection or connect an existing OpenSearch Serverless collection from the Vercel console. AWS also cited OpenSearch Agent Skills for Claude Code, Cursor and Kiro as another way developers can use coding agents to build search applications, work with logs or manage migration tasks.
The updated service is now generally available in AWS commercial Regions where OpenSearch Serverless is available, with compute and storage billed separately. The immediate update is faster and more elastic infrastructure. The question now is whether OpenSearch can make search and vector retrieval dependable enough for enterprises in this new agentic era.
