OpenSearch 3.8 Boosts AI Agents, Vector Search, and Observability

OpenSearch 3.8, released on August 4, 2026, marks a significant leap forward for the open-source search and analytics suite, with a sharp focus on AI agent infrastructure, vector search performance, and observability workflows. Authored by James McIntyre, the update introduces concrete engineering optimizations that deliver measurable gains in throughput, latency, and usability—addressing long-standing pain points for enterprise-scale AI and retrieval-augmented generation (RAG) workloads.

For organizations managing large-scale vector databases or agentic AI systems, the performance improvements are substantial and immediately actionable. The release prioritizes efficiency in high-dimensional data processing, a critical factor as AI applications demand faster, more scalable retrieval and inference. With these enhancements, OpenSearch 3.8 positions itself as a compelling alternative for teams seeking to reduce infrastructure costs without sacrificing performance.

Performance Breakthroughs in Vector and Radial Search

OpenSearch 3.8 delivers its most tangible advances in vector ingestion and radial search, where targeted optimizations yield double-digit percentage improvements in key metrics. One of the standout changes is the introduction of Base64-encoded ingestion for knn_vector fields. This reduces network payload by 74%, compressing a 16 KB JSON payload to just 4 KB for 768-dimensional floats. The impact is immediate: bulk ingestion throughput increases by up to 4.16x, while median latency drops by 83%. For applications ingesting high-dimensional embeddings at scale—such as those powering real-time recommendation systems or semantic search—this optimization translates directly into lower operational costs and faster data processing.

The update also overhauls radial search, a graph-based similarity query mechanism used in applications like fraud detection and knowledge graphs. A redesigned traversal engine now achieves up to 2.1x higher query throughput and 45% lower median latency on datasets containing 10 million vectors. Beyond speed, the quality of results has improved: mean recall for radial queries jumps from 0.85 to 0.97, representing a 14% gain in accuracy. For latency-sensitive use cases, the p90 latency for radial queries has plummeted by up to 77% compared to OpenSearch 3.7. This reduction in tail latency is particularly critical for user-facing applications, where even occasional delays can degrade the overall experience.

AI Agents and Standardized Tooling

The release significantly expands OpenSearch’s AI agent capabilities, with a particular emphasis on standardization and interoperability. A major highlight is the extension of Model Context Protocol (MCP) support to all four agent types, including flow and conversational flow agents. This enables OpenSearch agents to connect seamlessly to external MCP tool servers, fostering a more integrated and flexible ecosystem. A new list tools API further enhances this functionality by allowing programmatic discovery of available external tools, streamlining the development and deployment of agent-driven workflows. Abdul Muneer Kolarkunnu of NetApp Instaclustr contributed to these MCP upgrades, underscoring the collaborative nature of the project’s development.

Beyond MCP, OpenSearch 3.8 introduces gRPC transport for ML Commons, a move that addresses the growing demand for efficient, low-latency communication in AI workloads. The new PredictModelStream and ExecuteAgentStream APIs support streaming inference, reducing CPU overhead compared to traditional REST or Server-Sent Events (SSE) approaches. This is particularly beneficial for high-throughput inference scenarios, where minimizing computational overhead can lead to significant performance gains.

The update also broadens the scope of the LLM-as-a-Judge feature, which now supports a wider range of third-party large language models, including Azure, DeepSeek, Ollama, Google Gemini, Anthropic Claude, and Amazon Bedrock. This expansion, enabled via connector blueprints, allows organizations to leverage their preferred LLM providers for evaluation tasks, enhancing flexibility and reducing vendor lock-in.

To address the challenges of managing long-running agent deployments, OpenSearch 3.8 introduces an experimental agentic memory feature. This includes retention policies that automatically delete expired sessions on a 24-hour schedule, helping teams control storage growth and maintain system efficiency without manual intervention.

Observability and Query Experience

OpenSearch 3.8 takes direct aim at a persistent criticism of its Piped Processing Language (PPL): the lack of intuitive tooling and guidance for users. The release introduces a visual PPL builder and in-editor linting, providing real-time feedback and simplifying the process of constructing complex log analysis queries. This addresses prior industry feedback that PPL lagged behind competitors like Elastic’s ES|QL in terms of usability. Additionally, the inclusion of SQL support in Discover further lowers the barrier for operations teams adopting OpenSearch for observability tasks, allowing them to leverage familiar query syntax alongside PPL.

The update aligns with broader trends in the observability ecosystem. On July 28, 2026, the OpenTelemetry Demo 3.0 was released, introducing agentic tracing and renaming attributes (e.g., from app. to demo.) across 6,859 forks. This shift reflects a growing industry emphasis on standardized, agent-driven observability workflows—a trend that OpenSearch 3.8 is well-positioned to support with its enhanced PPL tooling and agent capabilities.

Competitive Context and Open Questions

OpenSearch 3.8 arrives in a competitive landscape where proprietary and open-source solutions vie for dominance in AI and observability. In a September 2025 analysis, Elastic—a direct competitor—criticized OpenSearch’s PPL for lacking intelligent autocomplete and cross-cluster support, features that Elastic’s ES|QL language offers. The visual PPL builder and SQL support in Discover introduced in this release appear to be a direct response to these criticisms, though it remains to be seen whether these changes will fully satisfy Elastic’s concerns or those of users accustomed to more mature query languages.

While the performance metrics in OpenSearch 3.8 are unambiguous, some questions remain unanswered. The release notes do not detail migration paths for non-Lucene backends, leaving uncertainty for teams using alternative storage engines. Additionally, the long-term adoption of features like agentic memory and MCP integration will depend on community feedback and real-world testing. Nevertheless, the concrete improvements in vector search, radial query performance, and agent tooling provide a strong foundation for organizations evaluating open-source alternatives for their AI and observability workloads.

For executives and technical specialists, OpenSearch 3.8 offers a compelling value proposition: measurable reductions in latency (up to 83% for vector ingestion and 77% for radial queries) and standardized agent tooling via MCP. These enhancements reduce the infrastructure cost of RAG and agentic AI workloads while improving the usability of complex query logic. The visual PPL builder, in particular, lowers the barrier for operations teams to adopt OpenSearch without relying on external expertise, making it a practical choice for a broader range of users.

Looking ahead, OpenSearch’s focus on AI agent infrastructure and performance optimization suggests a strategic commitment to serving as a cost-effective, high-performance platform for enterprise AI. The next challenge will be ensuring these improvements translate into widespread real-world adoption, particularly among teams that have historically relied on proprietary solutions. For now, OpenSearch 3.8 delivers the concrete performance gains and usability enhancements that professional users have been waiting for.