November 2026 Summaries

No summary generated yet.

LangChain’s internally developed GTM Agent automates the creation of sales collateral such as account plans, proof-of-value documents, proposal decks, and onboarding plans, generating 1,540 artifacts over 12 weeks and more than 100 per week. The system combines AI models with business-specific context, including Salesforce data, product usage, call transcripts, sales-stage requirements, templates, brand rules, and leadership feedback, rather than relying primarily on prompting. Artifacts are generated through a workflow that gathers missing information from reps, populates templates in a sandbox, renders and self-reviews files for layout and brand consistency, and delivers editable documents through shared account folders and Slack. Proactive functions, including pre-meeting briefs, task assignments, and an artifact hub, help reps identify needed materials before sales-process gaps emerge. LangChain reports that 47% of deals with a sent artifact progressed to a later stage versus 16% of matched deals without one, while customer calls followed 46% of generated artifacts within 14 days and externally shared materials saved an estimated 337 rep hours.

Workday developed an internal agent platform to help engineering teams rapidly build, deploy, and operate AI agents while meeting stringent enterprise requirements for security, auditing, compliance, and data governance. Initially built by four people, the platform combines LangGraph for flexible graph-based agent orchestration and LangSmith for tracing, observability, and evaluations, alongside a Workday-specific Python SDK that connects LangChain-style primitives to internal services such as LLM gateways, RAG retrieval, and prompt tooling. Its Agent System of Record governs the APIs agents can access, while standardized infrastructure supports production deployment, authentication, and protocols including A2A and MCP. The platform reduced prototype development timelines from four or five months to less than one day and has become a unified interface for previously isolated internal AI services, contributing to the team’s growth to roughly 30 members.

Eval-driven development (EDD) adapts test-driven development principles to AI agents by defining measurable success criteria before implementation, then using repeatable evaluations to compare changes and detect improvements or regressions. Because agents are nondeterministic, may take varied tool-use paths, and can produce multiple valid responses, evaluations can assess outputs, actions, tool calls, multi-step behavior, or resulting system state through deterministic code checks or LLM-as-a-judge methods. The process begins by specifying a narrowly scoped behavior and its requirements, creating a representative dataset from real or synthetic requests, defining an evaluator, and running a baseline before modifying the agent. A ticketing-tool example illustrates how comparing experiments can reveal hidden regressions, such as a broad tool description causing a VPN how-to question to create a ticket rather than search documentation, which tracing can diagnose and refine. The approach recommends held-out datasets to prevent overfitting, automated evaluations in CI/CD, and production monitoring through online evaluators, human annotation, dashboards, and issue clustering. Common pitfalls include vague success definitions, designing tests after a fix, relying only on average scores, adding unnecessary evaluations, and using LLM judges without validating them against human assessments.

Async subagents in Deep Agents are designed to let an orchestrator remain responsive to users while background workers handle lengthy tasks in parallel, addressing the limitations of traditional ReAct loops where tool calls and subagents block all other work. The orchestrator retains conversation context, launches tasks without waiting for their completion, and can monitor, update, cancel, or list them through dedicated task-management tools. Running subagents can be redirected either by interrupting and restarting their work for corrections or by queuing additional instructions for completion afterward. Task results are communicated automatically whether the orchestrator is active, idle, or awaiting user approval, while approval requests can be handled within a subagent’s own thread. Available in Python, JavaScript, and Managed Deep Agents, the feature reflects a broader industry shift toward asynchronous agent and tool execution, though task notifications remain in beta.

As AI agents handle longer, tool-using, multi-turn tasks, effective evaluation requires human review of more than final responses, using complete agent traces that record messages, model outputs, tool calls, and intermediate actions. The discussion outlines three principles for human-in-the-loop workflows: domain experts define quality through their judgments and explanations, reviewers need full conversational context to make reliable assessments, and structured feedback should become reusable evaluation data. LangSmith annotation queues support this process by allowing reviewers to assess individual runs or full conversation threads, with trajectories offering a simplified chronological view of messages and tool use. Human scores and comments can expose gaps in automated evaluators, refine rubrics, align LLM judges with expert standards, and create golden datasets for offline testing. Fetch Rewards illustrates this approach by having reviewers verify that its AI assistant’s answers are grounded in tool results, accurately represent offers and product images, and match users’ eligibility, with engineers using reviewer reasoning to improve automated judging.

Browser agents are increasingly used for multi-step tasks on websites without APIs and for end-to-end testing, but their effectiveness depends heavily on browser tool design because web tasks are long-running and generate large amounts of potentially irrelevant context. In tests across WebVoyager, HardBench, and WebTailBench, LangSmith’s Managed Deep Agents team compared seven browser toolsets and selected Browserbase’s Stagehand, which achieved the highest WebVoyager success rate at 90% while costing substantially less and operating faster than alternatives such as Playwright MCP. Stagehand’s advantages stem from code-based execution that batches many browser actions into one call, compact accessibility-tree snapshots that identify actionable page elements without raw HTML or image interpretation, and browser-local execution that reduces Chrome DevTools Protocol round trips. The team further improved performance by replacing fixed browser wait times with adaptive waits that finish once pages become idle, reducing median HardBench task time from 182 to 124 seconds without meaningful accuracy loss. The findings emphasize that tracing agent behavior and validating changes through evaluations can improve browser-agent efficiency, cost, and reliability, while Stagehand is available for both managed and open-source Deep Agents deployments.

LangSmith Bring Your Own Cloud (BYOC) is now generally available on AWS and in beta on Azure and Google Cloud, allowing organizations to run LangSmith’s sensitive data plane within their own cloud accounts while LangChain manages the control plane, orchestration, scaling, updates, and runtime infrastructure. The split architecture keeps traces, prompts, datasets, experiments, deployments, secrets, and model-provider integrations in customer-controlled environments, while LangChain retains authentication, organization configuration, billing metadata, hashed API keys, and UI assets. Organizations can operate multiple data planes across accounts, regions, environments, or business units, enabling teams to meet compliance, network, and data-residency requirements without the maintenance burden of self-hosting. BYOC uses cloud-native components such as private AKS networking on Azure, region-flexible namespace-scoped resources on Google Cloud, and customer VPCs, selected database subnets, and security groups on AWS. It supports core LangSmith capabilities, including tracing, evaluations, LLM Gateway, deployments, Engine, isolated sandboxes, and Managed Deep Agents endpoints, while offering migration tooling for existing self-hosted workspaces and AWS trace migration.

Although agent architectures have converged around LLMs operating in loops with tools and sandboxes, the rise of proactive, always-on, goal-oriented agents is creating new context-engineering challenges. Established techniques such as summarization, tool-result offloading, progressive-disclosure skills, tool search, subagents, and bash-based execution address many traditional agent limitations, but they do not fully solve issues introduced by persistent operation. Developers must determine how agents can be helpful without overwhelming users with notifications, what information to retain in memory, how to coordinate long-running asynchronous tasks while remaining responsive, and how to track and iteratively pursue broad goals over extended periods. Designing goals that cannot be gamed is also difficult, as agents may optimize objectives in unintended ways, while browser-based interaction introduces additional complexity around navigating visual and DOM-based information.