LangChain unveils SmithDB — distributed database for agent traces

LangChain has introduced a new piece of infrastructure for the fast-growing business of monitoring AI agents: its own database. The company said on May 13 that it launched SmithDB, a purpose-built distributed database for agent observability that now powers core workloads inside LangSmith, LangChain’s platform for tracing, evaluating and deploying agents. The launch was detailed in a company blog post by Ankush Gola and was also included in LangChain’s May 14 roundup of products announced at Interrupt, the company’s conference in San Francisco. The announcement gives a more concrete picture of how LangChain is responding to a problem many agent developers have run into: traces are getting larger, more nested and harder to query with conventional observability systems. LangChain said modern agent traces can contain hundreds of nested spans, multimodal payloads such as images and audio, and operations that remain open for hours before completion events arrive. (langchain.com) ### What exactly did LangChain launch? SmithDB is a distributed database built specifically for agent observability, according to LangChain. The company said the system now backs core LangSmith workloads and is intended to improve the speed of loading traces, filtering runs and searching across agent activity. Jacob Talbot, writing in LangChain’s Interrupt product roundup on May 14, described SmithDB as “the database purpose-built for agent observability.” That places it below the LangSmith product layer and squarely in the data infrastructure stack LangChain is assembling around production agents. (langchain.com) ### Why did LangChain say existing databases were not enough? (langchain.com) LangChain said the shape of agent data has changed since LangSmith launched in 2023. The company said early AI applications were closer to retrieval pipelines and prompt chains, while newer systems generate larger traces with more nested calls, longer runtimes and more varied content types. (langchain.com) Ankush Gola wrote that the resulting workloads require support for random access, interactive filtering, full-text search, JSON filtering, tree-aware queries, thread reconstruction and aggregations over large traces. LangChain said those requirements created query patterns that “general-purpose databases were never designed to handle.” ### How is SmithDB built? (langchain.com) LangChain said SmithDB is built in Rust and uses Apache DataFusion and the Vortex file format. The company said the architecture stores durable trace data in object storage, keeps a small Postgres metastore, and uses stateless ingestion, query and compaction services. Object storage is central to the design LangChain described. (langchain.com) The company said that approach lets SmithDB scale by adding compute instead of managing local disks, and makes the system easier to deploy in self-hosted and multi-cloud environments where customers may need tighter control over data location. ### What performance numbers did LangChain publish? (langchain.com) LangChain said SmithDB delivers P50 latencies of 92 milliseconds for trace-tree loads, 400 milliseconds for full-text search and 82 milliseconds for run filtering. In its May 13 post, the company said those gains make core LangSmith experiences up to 15 times faster than before. Those figures came from LangChain’s own announcement, and the company did not provide an independent benchmark in the materials reviewed here. (langchain.com) The numbers are still notable because they show which workloads LangChain sees as central to agent observability: loading a trace, filtering large datasets and searching inside inputs and outputs. ### Where does this fit in LangChain’s broader product push? Interrupt 2026 took place on May 13 and May 14 at The Midway in San Francisco, according to LangChain’s event preview. The company used the conference to unveil a wider set of products, including LangSmith Engine, Managed Deep Agents, Context Hub and other LangSmith features alongside SmithDB. (langchain.com) LangChain’s own materials frame SmithDB as part of a broader effort to support agents “at enterprise scale.” In the preview for Interrupt, the company said this year’s event focused on the tooling, infrastructure and team structures needed when agents move beyond proofs of concept. May 14 is the clearest next marker in the public record so far: LangChain’s post-event roundup links SmithDB alongside the rest of the Interrupt launches, and the company’s May 13 technical post says the database already backs core LangSmith workloads. (langchain.com 1) (langchain.com 2)