The latest AI news and updates of LangChain — AI company building frameworks and observability tools for developing and orchestrating LLM agents. Covering LangChain's latest product updates, launches, and company news from the past 90 days.
LangChainSep 10LangChain integrated the Harbor framework into Managed Deep Agents to provide reproducible evaluation environments. Each agent evaluation now runs in a fresh container, with results automatically tracked in LangSmith. This integration allows developers to verify that skill additions, model swaps, or tool description tweaks maintain agent performance without breaking existing functionality.
LangChainSep 9LangChain integrated Browserbase’s Stagehand SDK into its Deep Agents framework, providing browser tools without custom definitions. The integration ships as an MCP server offering run, snapshot, and screenshot capabilities. It requires a persistent ClientSession and uses stdio transport to isolate dependency environments, supporting agent interaction with web pages through Browserbase or local Chrome.
LangChainSep 9LangChain launched Connections in Managed Deep Agents 0.7.0+, moving credential management from local .env files to the LangSmith workspace. This update introduces agent-owned secrets for shared tool access and user-owned OAuth, which resolves credentials per caller at runtime. Agents now automatically handle OAuth flows and carry the identity of the specific user requesting the action.
LangChainSep 4LangChain analyzed LangSmith observability data from the 14 days ending August 30, 2026, to track model usage. gpt-4o-mini led in organizational adoption at 13%, while gpt-4.1-mini drove 7% of total call volume. DeepSeek V4 Flash ranked second in adoption at 9%, marking the only open-weight model to appear in the top-10 lists for reach or volume.
LangChainSep 3LangChain updated its Python MCP adapters to support the new stateless Model Context Protocol specification. The client now defaults to a stateless architecture, where each tool invocation creates a fresh session, while maintaining support for stateful sessions via explicit configuration. These updates also include improved error handling for tool execution failures.
LangChainAug 30LangChain published a new Max Agency podcast episode featuring Unify CTO Connor Heggie. The discussion details how Unify reduced AI agent compute costs by 90–95% two weeks before launch. Engineering fixes included optimizing prompt caching, replacing full virtual machines with a suspending Python REPL, and implementing subagents as function calls to prevent runaway budget depletion.
LangChainAug 26LangChain published a case study detailing Rippling's production AI evaluation pipeline, which uses four distinct layers to ensure system health. The architecture includes offline mock-based evals, post-merge integration tests, deploy-blocking scenarios, and continuous production monitoring. Rippling uses LangSmith to manage these traces and analyze conversation data at scale across its multi-agent system.
LangChainAug 26LangChain launched WikiBench, a benchmark for evaluating codebase documentation agents on the Harbor framework. Testing shows that combining a wiki with raw source code improves accuracy and cost-efficiency for coding agents compared to using source code alone. The benchmark uses a reader agent to score wiki quality based on grounded answers to repository-specific questions.
LangChainAug 25LangChain updated LangSmith Engine to deliver over twice the performance on internal benchmarks for identifying agent issues. The platform now supports self-hosted deployments and includes a Reduced Analysis mode for cost-sensitive users. Additionally, the tool adds Slack and Linear integrations for issue tracking and automatically closes stale issues that no longer appear in production traces.
LangChainAug 25Toyota North America scaled its enterprise AI portfolio from one production agent every six months to over 50 agents in four days. The team built this platform using LangChain’s Deep Agents and LangGraph, while utilizing LangSmith as an observability board to monitor agent performance, tool failures, and pipeline regressions across the manufacturing line.
LangChainAug 22LangChain launched LangSmith Tuned Evaluators to automatically score agent behavior in production, starting with Perceived Error. These managed evaluators replace manual prompt and infrastructure maintenance, delivering frontier-level accuracy at an 82% lower cost. The Perceived Error evaluator analyzes threads with at least two message pairs and is available now for Plus and Cloud Enterprise plans in the US.
LangChainAug 22LangChain launched LangSmith Preview Builds in public beta for LangSmith Deployment. This feature creates temporary, production-like environments from pull request branches, allowing teams to test agent changes, inspect traces, and share feedback before merging. Previews are short-lived and automatically tear down after the pull request merges, streamlining the agent review process without extra CI/CD setup.
LangChainAug 12LangChain launched LangSmith Bring Your Own Cloud on AWS, providing a managed deployment within a customer's own VPC. This model keeps sensitive agent traces, datasets, and runtime data inside the customer's AWS boundary while LangChain handles infrastructure provisioning, scaling, and upgrades. BYOC is available for Enterprise customers across 15 AWS regions in the US, EU, and APAC.
LangChainAug 12LangChain rebuilt LangSmith Dashboards to enhance investigation and reporting workflows. The update introduces free-form layout arrangement, KPI widgets alongside trend charts, and cross-unit metric comparisons. Dashboards now support trace breakdowns by model or user, custom notes, and configurable charts for tracking specific application metrics.
LangChainAug 11LangChain benchmarked NVIDIA’s Switchyard router on 145 multi-step agent tasks, finding 93% of turns could be handled by a 30B model instead of a frontier model. This routing strategy reduced total costs by approximately 70% while retaining 90% of frontier-level accuracy. LangChain released a Deep Agents integration to enable this routing in production agent workflows.
LangChainAug 8LangChain launched Managed Deep Agents in public beta, providing a hosted runtime for production-ready agents. The service handles durable execution, persistence, sandboxes, and memory, allowing teams to deploy agents via CLI without managing underlying infrastructure. It includes first-class support for Harbor evals and Slack channels, enabling focus on agent behavior and business logic.
LangChainAug 8LangChain’s LangSmith Signal report finds that one in three active teams ran an open-weight model in July 2026, up from one in five a year ago. Daily usage of open-weight models grew 9x year-over-year, outpacing the 4.8x growth seen in closed models. This analysis draws from production observability data across billions of agent runs on the LangSmith platform.
LangChainAug 6LangChain added Moonshot AI's Kimi K3 and K2.6 models to the LangSmith LLM Gateway. Powered by Fireworks AI, the integration allows model access using a LangSmith API key and Gateway Credits, eliminating the need for separate provider accounts. Usage is governed by existing workspace spend policies and includes full LangSmith tracing for production observability.
LangChainAug 5LangChain updated OpenWiki with a built-in web visualizer, accessible via the openwiki visualize command. The tool serves a local interface with an interactive node graph and a live Markdown reader for exploring repository documentation and file relationships. The visualizer automatically updates as wiki files change during the session.
LangChainAug 5LangChain released an autonomous SRE agent for Kubernetes that performs proactive health checks and on-demand diagnostics. The agent uses a scheduler to bypass full model calls for routine checks, reducing costs by 95–99%. It features a human-in-the-loop safety model for all write actions and uses LangSmith Engine to automatically detect and fix issues in production traces.
LangChainAug 4LangChain published a framework for evaluating voice agents across three distinct dimensions: Execution, Outcome, and Experience. The approach measures instruction adherence, goal achievement, and conversational smoothness separately to identify specific agent weaknesses. Evaluations use LangSmith to combine code evaluators, LLM judges, and business-system checks for monitoring performance and tracking improvements over time.
LangChainAug 3Stripe built Kai, its company-wide Knowledge AI Platform, in one week using LangChain’s Deep Agents harness. The agent connects to internal data, Slack, and Google Suite, reaching over 5,000 users in four weeks. It currently supports 1,000 skills and 500 internal MCP tools, with 83% of Stripe employees using it weekly for data analysis and document drafting.
LangChainJul 31LangChain released ReviewBench, a benchmark designed to evaluate code review performance based on real pull-request feedback. The system curates actual review issues into reproducible tasks that run on the Harbor framework. This provides a standardized way to measure how effectively AI agents identify and address the specific types of issues human reviewers catch in production code.
LangChainJul 30LangChain released the LangSmith LLM Gateway in public beta, providing a centralized governance layer for production AI agents. The gateway enforces spend and rate limits, configures model fallback policies, and redacts sensitive data like PII and secrets before requests reach model providers. It supports multiple providers and custom endpoints to help manage agent costs and reliability.
LangChainJul 29Similarweb evaluates its Deep Research agent using a workflow built on LangSmith. The system combines deterministic tool-call checks, rubric-scored LLM judges, and faithfulness verification against retrieved data. Connecting these scores to traces and A/B baseline comparisons makes agent performance inspectable, enabling evidence-based decisions instead of relying on single-output quality checks.
LangChainJul 29LangChain updated OpenWiki to connect directly to LangSmith tracing projects. By analyzing how coding agents like Claude Code and Codex interact with a repository, OpenWiki identifies where agents fail or lack context. It then uses these insights to generate more accurate and useful documentation for the codebase.
LangChainJul 28LangChain’s internal data agent now handles roughly 40x the request volume of its three-person data team. This shift redirects the team to focus on building the models, business context, and guardrails that ensure agent reliability. The company’s agent-first data stack integrates Hex, dbt, and semantic models to provide trusted, self-serve analysis across the organization.
LangChainJul 27LangChain’s LangSmith Insights feature now converts production agent traces into targeted test questions. CreditGenie used this capability to generate 100–200 test questions per behavior after debugging thousands of agent traces. This workflow automates the creation of evaluation datasets directly from real-world production data, helping teams identify and resolve agent failures more efficiently.
LangChainJul 27LangChain released Deep Agents v0.7.0b2, reducing default-agent input tokens by 65% compared to v0.6.12. The update introduces in-place middleware overrides, replacing built-in instances with custom settings without disabling defaults. Additionally, planning todos are now opt-in, requiring explicit restoration via middleware configuration.
LangChainJul 24Bridgewater Associates presented PAT, an internal AI analyst deployed to hundreds of investors that performs hours of deep exploratory research in minutes. The tool leverages proprietary data and methodologies, featuring per-user security, autonomous self-correction, and parallel code generation. PAT operates as a component of Bridgewater’s artificial investor, executing complex analytical investigations through a compiler-like agentic architecture.
LangChainJul 22LangChain launched the Eval Engineering Skill to help coding agents build evaluation tests. The skill generates evals by analyzing context from a code repository and agent execution traces. This provides a measurable way to assess agent performance and reliability using production data from the development lifecycle.
LangChainJul 22LangChain highlights a community-developed extension that adds native browser control to its dcode coding agent. The implementation uses the agent’s /goal command to define objectives for browser-based tasks. This capability allows the agent to interact with web interfaces directly, expanding its utility for complex, multi-step engineering workflows.
LangChainJul 22LinkedIn engineers detailed their hiring agent at LangChain’s Interrupt conference, which uses LangGraph to cut time-to-interview by 60% for small businesses. The agent employs a central planner with a plan-execute-replan loop and a context-driven human-in-the-loop design to manage stateless scalability and ensure consistent recruiter-facing action paths.
LangChainJul 21Apollo.io rearchitected its AI Assistant using LangChain’s Deep Agents, reducing dev-to-launch time for new skills by 80-85%. The new skill-based architecture replaces a rigid supervisor system with dynamic goal-based execution. Apollo also introduced AI Watchtower, a six-layer evaluation framework, and expanded the assistant into a headless agent accessible via API and MCP server.
LangChainJul 21LangChain launched LangSmith tracing for voice agents built with Pipecat, LiveKit, OpenAI Realtime, and Gemini Live. The integration captures conversation audio, speech-to-text and text-to-speech latency, voice activity detection events, and interruptions. This provides full observability into voice pipelines, allowing teams to debug errors and evaluate agent behavior in production using the same workflows as text-based agents.
LangChainJul 21LangChain released a tracing plugin for Cursor that captures agent sessions, including model runs, tool calls, subagent activity, and file attachments, directly into LangSmith. The plugin uses the coding-agent-v1 schema, providing a unified dashboard for monitoring and comparing agent performance across Cursor, Claude Code, and Codex.
LangChainJul 18LangChain released Open SWE, an open-source, cloud-hosted coding agent that autonomously plans, writes, tests, and reviews code. The system integrates directly with GitHub, allowing tasks to be triggered via issue labels. It uses a multi-agent architecture with dedicated planner and reviewer components to execute complex engineering tasks in isolated sandboxes without requiring local resources.
LangChainJul 16LangChain details three middleware features powering the Box Agent. Parallel citation generation streams answers without user interruption, while prompt caching reduces latency and costs in multi-turn conversations. Additionally, the system automatically summarizes conversation history exceeding 170,000 tokens to prevent context overflow, ensuring the agent maintains performance at enterprise scale.
LangChainJul 15LangChain published a conceptual guide detailing the requirements for governing AI agents in production environments. The guide advocates for isolated, hardware-virtualized sandboxes to ensure safe execution, credential management, and full auditability. It outlines four core primitives for agent environments, including microVM isolation and snapshot-based state persistence, to help teams manage access, protect data, and control costs.
LangChainJul 15LangChain now allows one-click deployment of Fleet agents into Slack. Each agent receives a custom identity, name, and icon, enabling it to participate in channels and threads. The integration supports file handoffs and human approval workflows directly within Slack, keeping project context and agent interactions in one place without requiring code.
LangChainJul 15Toyota’s enterprise AI team built ToyotaGPT, a platform powered by LangGraph that reduced agent delivery from six months to four days. The system supports over 50 agents in production, utilizing a unified tool layer and automated skill generation from unstructured data. This architecture deploys agents via configuration files, replacing manual engineering workflows.
LangChainJul 14LangChain updated LangSmith to provide full-session observability for Cursor, GitHub Copilot, Pi, and OpenCode coding agents. The platform captures the complete run tree, including model calls, tool usage, and subagent activity, with no extra instrumentation required. It also provides out-of-the-box token usage and cost tracking, allowing teams to query and compare agent performance across different coding tools.
LangChainJul 10LangChain built an agent that drafts a cited VC investment memo in about 90 seconds for $0.40 in API costs. The agent uses four parallel LangGraph nodes powered by the Perplexity Agent API to research financials, product, and market data. A tool-less synthesizer then compiles the findings into an auditable memo with primary source citations for every claim.
LangChainJul 10LangChain published a playbook detailing how to tune agent harnesses for NVIDIA’s Nemotron 3 Ultra. By optimizing system prompts and middleware, the agent achieved a 0.86 score on the Deep Agents suite, nearly matching Opus 4.8 at roughly 10x lower cost per run. This eval-driven approach improves agent performance without requiring model fine-tuning.
LangChainJul 10LangChain updated OpenWiki to support general-purpose memory via a new brains mode. This mode builds a local personal wiki from sources like Gmail, Notion, X, and web search, separate from the tool's existing code-documentation mode. Each mode uses distinct setup and update workflows to maintain agent-readable knowledge layers for different work contexts.
LangChainJul 9LangChain highlights Finch Legal, an AI-powered pre-litigation platform for personal injury firms that has grown 10x over the past year. Finch uses LangSmith as its production observability layer to trace, evaluate, and monitor quality and cost across workflows including client communications, medical-record follow-up, and document processing.
LangChainJul 9LangChain released a plugin that traces every Claude Code session directly into LangSmith. The setup requires three commands and one JSON block, taking approximately two minutes to complete. Once configured, the plugin captures all messages, tool calls, and subagent runs as inspectable traces, providing full observability for debugging agentic coding workflows.
LangChainJul 8LangChain and NVIDIA launched the NemoClaw Deep Agents Blueprint, an open reference architecture for building governed enterprise agent systems. The stack integrates Nemotron 3 Ultra, a tuned Deep Agents harness, and the OpenShell runtime. In evaluations, the blueprint achieved an aggregate score of 0.86 at a cost of $4.48 per run, roughly 10x lower than comparable models.
LangChainJul 7Schneider Electric runs over 60 AI agents in production across 100+ countries, all traced through self-hosted LangSmith. Their internal AI Assistant, One Jo, serves 160,000 employees. The company uses LangSmith’s observability, evaluation, and deployment pillars to manage these agents within strict data residency and cybersecurity controls.
LangChainJul 6LangChain introduces LangSmith Evaluation to assess AI agent performance using real production data. The platform identifies agent failures and quality issues by running evaluations before and after deployment. It supports human feedback, prompt optimization, and CI/CD integration to catch regressions and improve agent reliability throughout the development lifecycle.
LangChainJul 2LangChain updated LangSmith to provide unified tracing for Claude Code, Codex, Cursor, GitHub Copilot, Pi, and OpenCode. The platform normalizes logs from these tools into a consistent trace tree, metadata schema, and query syntax. This integration provides a single dashboard for monitoring and comparing agent activity and costs across multiple coding tools.
LangChainJul 1LangChain launched OpenWiki, a CLI tool that writes and maintains codebase documentation specifically for AI agents. It generates repository documentation, automatically updates it via GitHub Actions, and supports Q&A over the codebase. The tool also appends instructions to AGENTS.md and CLAUDE.md files to ensure coding agents reference the documentation during development.
LangChainJul 1LangChain added support for recursive language model (RLM) workflows to its Deep Agents framework. This inference strategy allows agents to recursively call themselves or sub-models to decompose complex tasks before finalizing an answer. The implementation uses CodeInterpreterMiddleware to improve reasoning accuracy on long-context problems, as demonstrated in performance comparisons against standard agents on the Ulong dataset.
LangChainJul 1Z.ai's GLM-5.2 model is now accessible in the dcode coding agent. Developers can connect the open-weights model in three steps: download dcode, select GLM-5.2, and add an API key. The integration brings frontier performance on open weights to agentic coding workflows without model hosting overhead.
LangChainJul 1LangChain reports that Pendo catches 60% of agent failures in its product agent, Novus, before they reach customers. The team uses the LangSmith trace dashboard to identify gaps between customer needs and agent performance, then creates new evaluation sets to resolve issues. This daily workflow turns production traces into proactive agent improvements.
LangChainJun 30LangChain introduced a code interpreter for Deep Agents that executes untrusted code using WebAssembly and QuickJS. This in-process isolation model constrains agent capabilities without requiring a full sandbox environment. The company also open-sourced the quickjs-rs runtime and langchain-quickjs middleware to enable secure, snapshot-based durable pauses for agent workflows.
LangChainJun 30LangChain integrated its Deep Agents, LangSmith Sandboxes, and LangSmith Observability into the Harbor evaluation framework. This unified stack runs agents in isolated, parallel cloud sandboxes while recording step-by-step traces and experiment results. The integration uses the --agent langgraph and --plugin langsmith flags to orchestrate agent trials with deterministic verification.
LangChainJun 25LangChain distinguishes between two agent patterns in its Fleet platform. General Purpose Chat handles ad-hoc, low-setup tasks that end when the answer arrives. Specialized Agents manage recurring work by providing durable instructions, scoped tools, persistent memory, and event-based triggers. This framework allows teams to graduate one-off tasks into repeatable, delegated responsibilities as patterns emerge.
LangChainJun 22LangChain released Deep Agents v0.6, introducing a code interpreter that lets agents execute tool calls programmatically within a runtime. By keeping intermediate results in the runtime state and returning only relevant output to the model, this feature reduces round trips and token consumption. It enables model-agnostic programmatic tool calling and recursive workflows for any agent.
LangChainJun 22LangChain identifies two conflicting requirements for production-grade agent sandboxes: the instant startup speed of serverless functions and the statefulness of full machines. LangSmith Sandboxes address this by using hardware-virtualized microVMs that allow agents to install dependencies, edit files, and persist state across sessions. This architecture provides isolated environments for executing untrusted, model-generated code without exposing host infrastructure.
Frequently asked questions