LangChain

Rippling runs agent evals in four tiers: offline fixtures with recorded mocks on each commit, 300-400 sandbox queries after merge, about ten deploy-blocking live scenarios, and scheduled production runs several times daily. LangChain's account is a short case study.

LangGraph 1.2.20 gives each interrupt() within a task its own id and adds a strict resume mode that refuses stale interrupt answers, tightening human-in-the-loop flows. The rest of the release is dependency maintenance and docs fixes.

📊We’ve rebuilt Dashboards in LangSmith to make them more useful for investigation and reporting. Put KPIs next to trends, compare metrics with different units, break traces down by model or user, add notes, and arrange the dashboard freely. https://t.co/A876TfbLUZ https://t.co/dVCQ8coVYw

LangChain has released Deep Agents Code (dcode), an open-source, model-agnostic coding agent, alongside a new LangChain Academy course covering installation and enterprise-oriented usage controls.

LangChain's October roundup covers what shipped at Interrupt: LangSmith AI Search, Engine v2 with red teaming, Trajectories, Fine-Tuning via Baseten or Fireworks, Managed Deep Agents v0.9, and FastMCP support in LangChain 1.4.0.

LangGraph 1.2.19 fixes streaming edge cases: trace_policy no longer rewrites node payloads, messages and tools events carry their graph namespace, paused tools are not reported failed, and ToolNode tool names appear in node metadata.

A LangChain guide examining how Lyft, Fastweb+Vodafone, and LATAM Airlines run customer-service AI agents in production, covering self-serve agent-building platforms, frontline copilots, semantic routing for ambiguous requests, and using evals as a shared language between engineers and domain teams to iteratively improve agent behavior.

LangChain's Managed Deep Agents 0.8 adds HTTP channels for any JSON webhook, user-scoped memory and credentials, Slack file transfer, and built-in Parallel web search. A walkthrough builds a receipt-logging agent reachable by text message.

LangSmith's new judge feature scores every production trace rather than a sample, checks more criteria per trace without cost scaling linearly, and can flag safety or security issues fast enough to trigger automated responses.

LangSmith is now available as a bring your own cloud deployment on @awscloud. Keep agent traces and runtime data inside your AWS boundary, while LangChain manages provisioning, upgrades, scaling, and support. BYOC gives teams data sovereignty with the speed of a managed service. https://t.co/YK4T7BDV2N

Claude Platform traces from Managed Agents and Messages API requests can be sent to LangSmith over OpenTelemetry, in public beta. Setup is a Console destination, or SDK spans via an OTel exporter, with a metadata-only or full-content mode.

LangChain compares TypeSafe AI's Jev, a typed non-generative 'System One' model, with GPT-5.6 and Claude Sonnet judges on a replayed Deep Agents weather task. Jev showed far lower score variance and cost, though the five-case test is narrow and the authors say so.

LangChain ran NVIDIA's Switchyard router on 145 Deep Agents tasks. Only 7% of turns escalated to Opus 4.8, yet they carried 68% of spend. Routing cost 74% less at 80% versus 86% accuracy. It adds a break-even formula, run-to-run cost variance and the judge model's 21% overhead.

LangChain revamped its MCP integration: support for the new stateless MCP spec, the integration folded into the main langchain package, elicitation delivered through interrupts, cacheable list endpoints, and a new adapter built on FastMCP. Expect import path and adapter changes.

LangChain reports Nemotron 3.5 Super VL, NVIDIA's open-weight 120B model (12B active), averages 87% on its Deep Agents behavioral suite and finishes Harbor-lite tasks in 55% less time than DeepSeek V4.1 Flash. Integration docs show setup via ChatNVIDIA with thinking mode and image input.

LangGraph 1.2.18 snapshots DeltaChannel every 100 updates instead of 1000, fixes error_handler recovery on all paths, streams the last checkpoint of runs stopped by interrupt_after or the recursion limit, and stops duplicating tags across subgraphs.

LangChain releases Tuned Evaluators in LangSmith, starting with Perceived Error, which flags production threads where an agent erred or misunderstood. A LangChain-trained model runs the judging, which the company says beats frontier models at up to 82% lower cost.

Prime Intellect's Will Brown argues continual learning will become routine software lifecycle work. He covers evals as production metrics, choosing between SFT, on-policy distillation and RL, using RL for harness drift and routing, and when specialized models beat frontier ones.

LangChain is turning decision models into a first-class LangSmith Gateway feature, starting with SemIf, an open-source decision model, free for a week and drop-in compatible with the TypeSafe SDK.

LangChain and Nevermined published a cookbook for letting an agent transact mid-task without a human in the loop. It covers delegating payment credentials under explicit limits, having the agent buy credits and top itself up to pay providers, and tracing each purchase in LangSmith.

LangChain added browser use to Managed Deep Agents. A root browser declaration sets session scope and idle timeout, and built-in tools cover snapshots, actions, screenshots and handing control to a person for login. Docs warn that user scope shares logins across threads.

LangChain's postmortem on Content Studio, a Slack design agent that makes on-brand videos, graphics and decks. Covers router skills, wiki context, code-enforced gates, async subagents, per-thread sandboxes, two-loop memory, and pruning rules as models improve.

LangChain and AWS shipped AgentCore Payments middleware. When a tool call hits a paid API returning 402, the middleware checks the price against a session budget, signs the payment and retries, and every transaction is traced beside the reasoning that caused it.

LangChain positions its Deep Agents framework around built-in context engineering: filesystems, subagents, and skills wired up specifically to keep long-running agent tasks from overloading the context window.