LangChain has added AI Search to LangSmith, its platform for monitoring and evaluating AI agents, so that teams can find agent sessions by describing a behavior in plain language instead of building filters or evaluators first. The feature was announced on October 13, 2026 and is in public beta on LangSmith Cloud at no extra cost.
The problem it targets is familiar to anyone running agents in production. Traces record every user message, model call, tool call and result, but questions such as whether personal data leaked, whether a user gave up in frustration or whether an agent claimed to finish work it never did are hard to answer with tags, metadata or keyword search. Until now, LangChain says, answering them meant writing an evaluator in advance or reading traces by hand.
LangChain chose not to rely on vector search for this. The company argues that squeezing a long trace into a single embedding loses the details a query depends on. AI Search instead runs a fast decision model, a small model trained to answer one narrow question cheaply, over each candidate session. Ordinary structured filters, such as errored runs or a given tag, narrow the pool first, and matches stream back while the scan continues.
The work happens on trajectories rather than raw traces. LangSmith's trace database, SmithDB, converts incoming traces from LangGraph, LangChain, OpenTelemetry, the Vercel AI SDK, the Anthropic and OpenAI SDKs and other sources into one session-level view that lists each message, tool call and result once. Because agents usually resend the whole conversation on every model call, removing those repeats shrinks the input and lets a single query work across agents built with different frameworks.
Sessions can still be enormous. LangChain sampled 150 conversations from Open SWE, its in-house open-source coding agent, and found some ran past a million tokens; in the median session, only about 2,000 of 72,000 tokens came from the user and the agent's replies, while roughly 85% was tool output. AI Search therefore fills a fixed budget of 8,000 tokens per trajectory, putting user messages and agent replies first and trimming the rest.
That design has trade-offs the company acknowledges. Questions whose evidence sits in tool output, such as an agent ignoring a tool error, may be missed under the current budget, and subtle distinctions, like a user leaving satisfied versus giving up, remain hard to classify. LangChain says it is exploring question-specific context selection, uncertainty signals and grouping of similar results.
Organisation admins can enable AI Search in settings, after which it appears on the tracing page; the beta covers the GCP US, EU and APAC regions. LangChain has not published accuracy figures, so how well the decision model separates true matches from near misses will only become clear through use.