September 2026 Summaries

TypeSafe AI’s Jev is presented as a decision model that returns structured, probabilistic answers rather than generating text, targeting narrow classification and routing tasks where it claims substantial speed and cost advantages over frontier LLMs. The approach supports “AI-powered software,” in which deterministic code controls workflows while models provide semantic judgment only at uncertain decision points. LangChain positions LangGraph as the orchestration layer for such systems, using nodes, state, and edges to manage context, route execution, compose workflows, and provide durable execution, human review, and observability through LangSmith. Examples include legal document review, where Jev classifies responsiveness, personal information, and potential privilege before escalating pages to an LLM or attorney when needed, and browser automation, where Jev selects among known interactive elements and falls back to an LLM below a confidence threshold. The broader argument is that AI systems will increasingly unbundle capabilities, using inexpensive specialized models for bounded decisions and reserving frontier LLMs for open-ended reasoning or low-confidence cases.

LangSmith announced a set of releases at Interrupt NYC aimed at helping teams build, deploy, govern, and improve domain-specific AI agents through enhanced runtime infrastructure, observability, evaluations, and an intelligence layer that uses production data. LangSmith Engine v2 adds proactive red teaming, broader detection of reliability, latency, cost, and inefficient-tool-use issues, plus automated validation of candidate fixes before users create deployment pull requests. Managed Deep Agents v0.8 introduces identity-scoped user memory, separated access controls, agent- and user-owned external credentials, expanded Slack file sharing, HTTP webhook channels, and built-in web search powered by Parallel. LangSmith Trajectories provides a chronological, readable view of agent sessions across agents, tools, and subagents to simplify debugging, expert review, annotation, and online evaluation. LangSmith Fine-Tuning and its smithtune CLI support supervised fine-tuning of open models using trajectory-derived datasets, managed training, evaluation, serving, and application integration, while Custom Apps enables teams to create and publish workspace-specific interfaces for reviewing traces, experiments, annotations, and other agent data.

LangSmith has introduced Custom Apps, a feature that enables AI engineering teams to build, publish, and run tailored user interfaces using LangSmith data directly within their workspace. Available through chat-based app generation for Plus and Enterprise users or through code templates and the LangSmith API, the feature removes the need for teams to independently manage hosting, authentication, permissions, and sharing. Custom Apps can support recurring workflows such as giving human annotators role-appropriate context, creating specialized views for comparing agent experiments and detecting regressions, and focusing trace reviews on criteria relevant to particular agent types. Published apps remain connected to LangSmith data and can be shared with teammates alongside traces, evaluations, annotations, and experiment results. Plus plans include one Custom App per organization, while Enterprise plans provide unlimited apps; both tiers include the chat experience.

LangSmith has introduced Trajectories, a chronological conversational view that simplifies analysis of long-running agent sessions by aggregating human, AI, and tool messages across main agents and subagents into a single ordered sequence. While traces remain the detailed source of truth for nested runs, retries, timing, metadata, and tool inputs and outputs, trajectories help users first identify the behavioral step where an agent went wrong before drilling into underlying execution data. The feature works with LangChain, LangGraph, Deep Agents, major agent SDKs, and coding agents, and supports debugging, subject-matter-expert review through annotation queues, and online evaluation of complete agent paths. Teams can filter poorly scored sessions for review, use feedback to improve prompts and agent logic, and save high-quality trajectories as datasets or exports for supervised fine-tuning, making production behavior more usable for agent improvement. Trajectories are available on all LangSmith plans in the United States.

LangSmith Engine v2 expands the platform’s in-product agent for diagnosing and fixing AI agent issues, building on its analysis of more than 70 million traces and tens of thousands of identified problems since its May launch. It scans production traces to group related failures, identify root causes, propose fixes, generate evaluation examples, and monitor outcomes, while new red-teaming capabilities proactively test agents for weaknesses such as hallucinations and system-prompt violations before they affect users. The release also identifies less obvious production problems, including inefficient tool-use trajectories and negative trends in error rates, latency, and cost. For LangSmith Deployment users in private beta, Engine can reproduce failures, iteratively test and refine proposed fixes, and provide validated changes that can be opened as pull requests for deployment. LangSmith reports that recent improvements have more than doubled issue-detection performance and improved fix effectiveness by 25%, while upcoming self-hosted support will add bring-your-own-key inference to keep trace data within customers’ virtual private clouds.

Managed Deep Agents 0.8 expands LangSmith’s production agent platform with user-owned credentials, user-level memory, HTTP channels, Slack file transfers, and built-in web search powered by Parallel. The code-first framework packages the Deep Agents harness with managed infrastructure for authentication, permissions, memory, tools, channels, scheduling, and evaluations, aiming to reduce the operational work required to deploy agents at scale. Its new layered memory model separates shared agent memory from caller-specific user memory, with access defaults designed to prevent personal information from appearing in group Slack conversations or HTTP interactions. The release also supports both agent-owned and user-owned connections to external services, enabling shared capabilities such as web search alongside person-specific access to systems such as GitHub, Linear, Notion, Salesforce, and Google Workspace. Agents can now receive files through Slack and connect to any webhook-capable service through HTTP, while Parallel search can be enabled as a managed MCP tool without separate credentials or API setup.

LangSmith Fine-Tuning and its smithtune CLI are introduced as a public-beta workflow for converting LangSmith agent trajectories into specialized supervised fine-tuned models, with managed training through Fireworks or Baseten and evaluation in LangSmith. Agent trajectories capture the full sequence of messages, tool calls, results, and changing tool context needed to reproduce successful behavior, allowing teams to curate “golden” examples, create auditable datasets, filter for model compatibility, and split data for training and evaluation. The CLI supports planning training configurations, LoRA-based managed training without GPU administration, validation-based checkpoint selection, replay evaluations comparing tuned and base models, and deployment of successful models. Internal experiments reported that fine-tuning Kimi K3 improved an issue-detection benchmark score from 90.0 to 96.0, while fine-tuned Qwen-3.8-27B improved code-review F1 from 48.9% to 53.7% while reducing model calls and tool requests by about 30%. The announcement emphasizes that careful data selection and review are crucial, notes that an earlier less-selective dataset reduced performance, and recommends fine-tuning primarily for recurring tasks where teams have clear examples of desired agent behavior after harness-engineering improvements have been explored.

Healthcare AI developers Abridge and Included Health use LangSmith to turn scarce clinical review into reusable evaluation infrastructure, including labeled datasets, calibrated automated judges, annotation workflows, and release gates. Included Health’s Dot care-navigation agent is evaluated for appropriate routing, use of conversational context, and emergency detection, while Abridge’s clinical-note system is assessed for accurate attribution, completeness, and avoidance of unsupported details that could affect medical records or billing. Because clinical ground truth can vary and correct inaction is often important, both organizations combine expert judgment with evaluation methods tailored to specific failure modes, specialties, and source conversations. Production reviews generate structured feedback that informs dashboards, product changes, and future tests, while staged releases, backtesting, A/B testing, and continuous monitoring help teams deploy changes more quickly without relying on new manual review for every iteration. Included Health reports increased engagement, clinician agreement above its target for care recommendations, and high detection of high-risk situations, while Abridge reports shortening release cycles from months to days. The approach also requires careful handling of protected health information through de-identification, access controls, auditability, and appropriate cloud or self-hosted deployment choices.

Jev, a System One model from TypeSafe AI, is now available in LangSmith as a judge for evaluating agent behavior through fast, typed decisions rather than generated text. Unlike deterministic code-based evaluators and slower, less consistent LLM-as-a-judge approaches, Jev assesses states such as agent traces using yes/no probability, multiple-choice, and ordered-score questions, making it suited to high-volume, narrowly defined criteria such as PII leakage, user intent, and frustration. TypeSafe claims Jev is substantially cheaper and faster than comparable LLM classification methods and can evaluate multiple criteria in parallel; a cited test against GPT-5.6 variants and Claude Sonnet 4.6 found that Jev matched human decisions, had much lower variance, averaged 0.44 seconds per call, and cost $0.00035 per call. LangSmith users can configure Jev by adding a TypeSafe API key, selecting the TypeSafe provider and jev-latest model in an evaluator, defining the evaluated state, and adding typed feedback questions, after which results can be filtered, charted, or used to trigger alerts and automations.

TypeSafe AI’s Jev is presented as a “System One” decision model that could offer a third approach to agent evaluation alongside deterministic code-based checks and slower, less consistent LLM-as-judge methods. Rather than generating text token by token, Jev evaluates structured state through typed Choice, Score, and Noul questions, returning answers, probabilities, and confidence values. In an experiment using five fixed weather-agent traces, Jev was compared with GPT-5.6 Luna, GPT-5.6 Terra, and Claude Sonnet 4.6 across 100 repetitions per case, with human labels used as the accuracy reference. Jev matched the human binary pass/fail labels on all 500 repeated decisions, showed substantially lower observed score variance than the LLM judges, and cost $0.00035 per call, suggesting potential for broader and more frequent offline and production evaluation. The results are limited to a small, bounded test and do not establish why Jev was more stable or whether its performance generalizes, while low-cost evaluators could also amplify incorrect judgments at scale; therefore, human review and alignment checks remain important.

Jev, developed by TypeSafe AI, is a non-generative “System One” model designed to make fast, structured decisions within AI agent workflows, where repeated LLM calls can otherwise increase latency and cost. Rather than producing text, it evaluates a provided state and returns typed probabilistic answers to choice, scoring, or yes-or-no questions, with multiple questions evaluated in parallel; TypeSafe AI reports substantially faster and less expensive classification performance than comparable LLM approaches. LangChain supports Jev through its TypeSafeClassifier integration, allowing developers to submit text, structured data, or messages and receive classification results that can guide agent behavior. Suggested applications include routing requests to appropriately capable models based on task complexity and using Auto Mode middleware to assess potentially risky tool calls before execution, while leaving open-ended reasoning and text generation to conventional LLMs.

Included Health uses Dot, an AI-powered healthcare guide built with LangGraph and Deep Agents, to help employer and health-plan members navigate medical, financial, and administrative needs across services such as virtual care, behavioral health, specialty care, and scheduling. The platform replaces rigid decision-tree approaches with conversational agents that can interpret individual context, identify emergencies early, ask relevant follow-up questions, and recommend appropriate care while maintaining personalization at scale. Its federated “supergraph” architecture allows separate product teams to own specialized workflows while shared prompts, skills, context management, and a virtual filesystem preserve consistent tone and prevent members from repeating information when moving between services. Skills act as a registry of care capabilities, eligibility rules, and edge cases, while clinical reviewers continuously assess routing recommendations and refine the system. Human advocates can take over uncertain conversations through durable execution, after which the agent resumes with the full interaction history, and the company is exploring parallel agent support for advocates. LangSmith supports clinical review, operational metrics, and simulations used to validate system changes; following its August launch, Dot reportedly produced a 75% increase in chat engagement, clinical recommendation agreement above its 95% target, and detection of more than 99% of high-risk situations.

Eroom’s law describes the rising cost of pharmaceutical R&D, with the cost of developing a new drug reportedly doubling every nine years despite broader technological advances, driven by the complexity of preclinical research, clinical trials, regulatory compliance, and scientific data review. LangChain presents Deep Life Sci as an open-source, agentic assistant designed for clinical and laboratory scientists, arguing that general-purpose AI lacks the specialized knowledge, integrations, customization, and auditability required in life sciences. Built on the Deep Agents harness, the tool can search major biomedical sources including ClinicalTrials.gov, PubMed, and PubMed Central; analyze uploaded scientific files; delegate work to sub-agents; and run data analysis code in sandboxed environments. Suggested uses include interpreting RNA-seq or proteomics screens, systematically identifying and extracting evidence from clinical trials, and helping prepare, review, and revise clinical documentation. Organizations can adapt the open-source system with internal data, preferred models, guardrails, and approval processes, while LangSmith provides end-to-end traces and evaluation tools to support auditing, debugging, monitoring, and iterative improvement.

LangChain describes building and open-sourcing a Slack-based Paid Media Agent to help a small marketing team scale from organic growth to five paid advertising channels, consolidating campaign, lead, and pipeline analysis while proposing human-approved optimizations. Within six months, paid media grew from zero to 20% of marketing pipeline, qualified-lead costs fell 30% despite a 60% spend increase, and internal reporting saved about $5,000 monthly; moving deterministic calculations from the model into code also made reporting roughly 40 times cheaper and 13 times faster. Built with LangChain Deep Agents, isolated LangSmith sandbox environments, structured skills, a company-specific wiki, live data tools, and reproducible code, the agent generates weekly platform reports, answers Slack follow-up questions, and recommends changes such as keywords, targeting, copy, and new campaigns. The development emphasized using models for judgment rather than calculations, assigning authoritative data sources by metric, discovering tools on demand to reduce context costs, isolating subagent state and permissions, and requiring human approval and post-change verification for campaign edits. LangChain plans to make the system more proactive and connect its marketing insights with broader go-to-market workflows.

Healthcare and life sciences organizations are developing AI agent systems under stringent patient-safety, privacy, compliance, auditability, and data-governance requirements, making observability, evaluation, and cost controls essential before granting agents greater autonomy. Common approaches include central platforms that consolidate fragmented pilots, agents for document-heavy back-office workflows with measurable returns, patient-facing conversational and voice assistants, clinician-support tools that retain human oversight, and longer-running scientific R&D systems requiring trajectory-level evaluation. Madrigal Pharmaceuticals created a modular multi-agent platform that normalizes disparate data sources, uses specialized skills and parallel sub-agents, and improves through production tracing and failure-driven evaluations. Abridge applies AI to clinical documentation and built a unified evaluation system that translates clinician feedback into measurable quality and safety criteria, enabling faster releases while testing agents across planning, tool use, boundaries, and clinical accuracy. Vizient uses a hierarchical agent architecture to help health systems query siloed operational and clinical data, with tracing and prompt versioning supporting reliability and maintainability. Across these examples, robust infrastructure for secure deployment, data governance, traceability, evaluations, and human review is presented as central to scaling trustworthy agents in high-stakes healthcare settings.

Credit Genie uses LangChain’s OpenWiki to automate repository documentation for its growing AI and ML engineering codebase, addressing stale manuals, fragmented Notion pages and READMEs, and knowledge concentrated among individual contributors. The company built a GitHub Pages portal that consolidates documentation from onboarded repositories into a searchable interface for employees, while coding agents are directed to repository-level openwiki folders to obtain current system context before making changes. OpenWiki detects meaningful code changes nightly, creates documentation-update pull requests, and is supported by automated approval, merging, and daily portal rebuilds to minimize manual maintenance. Credit Genie reports faster access to cross-repository knowledge, improved onboarding, reduced reliance on subject-matter experts, and more informed conversations among technical and nontechnical stakeholders. The next phase aims to integrate the company’s cross-repository dependency graph so documentation can explain system relationships and help engineers and agents assess the potential impact of code changes.

LangSmith Connections in the Managed Deep Agents prerelease provides named, runtime-resolved credentials that let agents distinguish between permissions and the identity on whose behalf an action is performed, avoiding hard-coded shared API keys and service-account attribution. Connections combine independent ownership and credential-type choices: agent-owned or user-owned credentials can each use static secrets or OAuth grants, with agent-owned secrets suited to shared services such as Tavily web search and user-owned OAuth enabling per-caller access to services such as GitHub, where actions and visible data reflect each user’s permissions and identity. Developers create connections with CLI commands, reference them in tools through `connections.get()` using a slug and owner type, and can rotate stored secrets without rebuilding deployments. OAuth support includes a catalog of providers requiring developers’ own app credentials, as well as MCP servers that self-register OAuth clients and expose tools directly, such as Linear. When users lack required grants, agent runs pause before model execution and present a consolidated authorization step, while token storage, refresh handling, callback routes, and consent-screen implementation are managed automatically. Local development supports agent credentials through environment variables and real user authorization under `mda dev`, while additional options support shared deployment OAuth accounts, scope restrictions, and custom OAuth providers.

Deepagents introduces context modes for supervisor-managed subagents, addressing the trade-off between context isolation and reuse in multi-agent workflows. Traditional isolated subagents receive only a task description in a new context window, which supports independent judgment and prevents unnecessary information from entering the supervisor’s context but can require repeated investigation. Forked subagents instead inherit the supervisor’s conversation history, allowing them to continue work with prior evidence and decisions while benefiting from prompt caching and reduced tool use. The framework recommends fork mode for workers and memory agents that depend on existing conversation context, while isolated mode is better suited to reviewers and researchers that need to assess work independently or investigate self-contained questions. Context modes can be combined with specialized tools, prompts, and permissions to tailor subagents for implementation, verification, research, or long-term memory tasks.

Nevermined’s cookbook describes integrating its agentic payment infrastructure with LangChain and LangSmith so AI agents can autonomously purchase metered tools and services while operating within predefined credit-card spending limits. It presents a research-agent example in which the agent buys Exa web-search credits after exhausting a free tier, automatically tops up when credits run out, and discovers and purchases Baselayer corporate-data access without prior integration, using a single delegated card authorization and x402-based payments. The approach emphasizes server-enforced budgets, time limits, per-transaction caps, and transaction-count controls intended to prevent uncontrolled spending, while keeping card details with Stripe rather than the agent or Nevermined. On the provider side, APIs, MCP tools, and agents can be protected and monetized through payment SDK integrations, allowing services to charge per call and record verification and settlement events in LangSmith traces. The broader argument is that usage-based AI services and small agent-mediated transactions are increasing, creating demand for payment systems that enable agents to procure capabilities in real time while making spending observable, reviewable, and tied to agent performance.

LangChain has revamped its MCP integration in response to rapid adoption of the Model Context Protocol and its major July 2026 specification update, moving support into the main `langchain.mcp` package and building it on FastMCP. The revised MCP protocol replaces session-based communication with a stateless core, improving server scalability and reliability by eliminating session pinning while enabling server-directed tool-list caching and elicitation, where tools can pause for user input or approval. LangChain represents elicitation through LangGraph interrupts, allowing an agent run to pause and resume after a reviewer supplies an answer, while cached tool catalogs can reduce repeated discovery requests according to server-defined TTLs. FastMCP supplies transport handling, authentication, protocol negotiation, logging, progress notifications, and compatibility with both legacy and current MCP servers, including mixed-server configurations through separate clients and automatically prefixed tool names. The beta Python implementation is available with `langchain[mcp]` version 1.4.0 or later, replaces the former `langchain-mcp-adapters` package and its multi-server client, and is expected to gain TypeScript support later.

Organizations across sectors are shifting from isolated chatbot pilots toward centralized agent platforms that provide shared infrastructure, governance, and production support, particularly for regulated back-office, compliance, security, and operational workflows where return on investment can be measured. Reported priorities include observability, evaluations, cost controls, data residency, and mechanisms for enabling business users while maintaining engineering guardrails. Schneider Electric supports more than 60 agents through a 350-person AI Hub, using traceability, evaluation frameworks, expert review, and dedicated deployment stacks to improve quality and manage critical-infrastructure requirements. Vodafone uses monitored, multi-step assistants to help engineers analyze network metrics and retrieve technical documentation, with end-to-end tracing used to locate failures in intent classification, SQL generation, or information retrieval. monday.com redesigned its Sidekick assistant from a single tool-heavy agent into an architecture of permission-aware orchestration, specialized subagents, controlled tools, and sandboxes, finding that clearer boundaries improved reliability, testing, and diagnosis. Across these examples, the central conclusion is that scaling agents depends less on deploying individual models than on building the surrounding platform, controls, feedback loops, and organizational practices needed to make their behavior measurable and continuously improvable.