LangChain AI Updates: November 4, 2026

1. LangSmith BYOC Reaches Azure and Google Cloud in Beta, Keeping Traces and Agent Deployments in the Customer’s Account

LangChain expanded LangSmith Bring Your Own Cloud (BYOC) on November 3: it is now generally available on AWS, where it launched in August, and in beta on Azure and Google Cloud. LangChain runs the control plane, which holds authentication, organization and workspace configuration, billing metadata, hashed API keys and static UI assets. The data plane lives in the customer’s cloud account and holds traces, prompts, datasets, experiments, agent deployments and workspace secrets, but LangChain still operates it, including scaling, schema updates and platform upgrades. An organization can run several data planes split by region, environment or business unit, even on different clouds, and each one runs tracing and evals, the LLM Gateway with the customer’s own provider keys, LangSmith Deployment, Engine and Managed Deep Agents endpoints. On Azure the data plane sits in a dedicated virtual network with a private AKS API endpoint, Istio and KEDA deployed through Helm, and system telemetry sent to Azure Monitor; LangSmith Sandboxes are not yet supported there. Google Cloud data planes can run in any region but are managed from LangSmith’s US control plane. On AWS the data plane runs inside an existing VPC, with the database in subnets the customer chooses and the customer’s own security groups attached to each application role. Self-hosted customers migrate in two steps: they provision a data plane, then run an open-source tool that copies users and roles, datasets, experiments, prompts, annotation queue configuration, automation rules and dashboards, with no downtime or changes on the source instance. Traces, runs and feedback are moved separately by LangChain, so far only on AWS. Terraform modules are public at github.com/langchain-ai/terraform, and data planes are provisioned through sales. Source

2. LangChain Publishes the Benchmarks Behind Its Deep Agents Browser: Stagehand’s Code Mode Ran at a Thirteenth of Playwright MCP’s Cost

LangChain published the evaluation behind the browser in Managed Deep Agents on November 3. It compared seven toolsets: native one-action-per-call tools (Playwright MCP, browser-use and Browserbase MCP), CLIs (agent-browser and Browserbase’s browse CLI) and code mode (Stagehand, and Browserbase MCP called from code). On 100 WebVoyager tasks with Claude Sonnet 5.5, pass rates ranged from 78% to 90%, but costs were far apart. Stagehand passed 90% for $1.98 in total at a median of 17.4 seconds per task, while Playwright MCP passed 78% for $26.71, about 13 times the cost, and took three times as long per task. On the 38 core tasks of Browserbase’s harder HardBench, run mostly with Claude Opus 5.5, Deep Agents passed 30 with Stagehand’s code-writing run tool and 21 with the browse CLI, at three times the cost per task. Code mode explains much of the gap. run averaged 10 steps per task against 26 for Playwright MCP, and a navigation returned a median of 117 characters against about 17,000 from Browserbase’s hosted navigate tool. Just calling Browserbase’s MCP tools from code instead of one at a time cut their cost from $17.87 to $4.52. Two other findings carry over to tool design in general. A version of run that took Stagehand’s native API instead of Playwright-style code passed 5.3 points fewer tasks and cost 19% more per passed task, because models write Playwright best. Capping accessibility-tree snapshots at 8,000 characters after the fact saved nothing, because agents simply fetched the page again. Traces also showed agents writing fixed sleeps such as page.waitForTimeout(8000), so run now ends a sleep once the page has been quiet for 500 ms. Combined with structured answers and a model-call timeout, that change cut median HardBench agent time from 182 to 124 seconds with accuracy within noise, and Stagehand has upstreamed it. Open-source Deep Agents users get the same tools through StagehandMiddleware in langchain-stagehand. Source

3. langchain-typesafe 0.0.1 Adds a TypeSafe Classifier Runnable and Experimental Model-Routing and Tool-Blocking Middleware

LangChain released langchain-typesafe 0.0.1 (MIT), the first stable version of its integration with TypeSafe AI’s decision models, following earlier alpha releases. TypeSafeClassifier is a LangChain Runnable that takes a state and a set of named questions and returns probabilistic answers. The state can be text or JSON, with LangChain messages allowed anywhere inside it. Questions come in three types: Choice (pick among criteria), Noul (true or false) and Score (a point on an ordered scale). The classifier needs a TYPESAFE_API_KEY. An [experimental] extra adds two agent middlewares whose APIs may change without notice. ModelRouterMiddleware classifies the latest human message once per agent run and sends the run to the model whose criteria match, for example a small model for well-scoped tasks and a larger one for deeper reasoning, and stores the answer with its probabilities in agent state so traces show why a model was picked. AutoModeMiddleware checks calls to explicitly listed tools against risky and safe criteria. When a call’s risk probability meets the threshold, it returns an error ToolMessage instead of running the call. Source