March 2026 Summaries

A collaboration between MongoDB and LangChain has resulted in a comprehensive platform that integrates AI agent capabilities directly into MongoDB Atlas, allowing teams to transition from prototype to production without rearchitecting their data infrastructure. This integration enhances agent functionality with features such as vector search, persistent memory, natural-language querying, and full-stack observability, enabling reliable deployment across the full pipeline. The platform supports seamless integration with LangSmith for end-to-end tracing and evaluation, allowing teams to maintain a single set of access controls while avoiding additional infrastructure. Clients like Kai Security and Fortune 500 enterprises are already leveraging these capabilities to enhance workflows in areas such as cybersecurity, compliance, and customer experience. LangChain's open-source frameworks, combined with MongoDB's robust data management, offer a scalable solution for deploying AI agents on existing infrastructure, facilitating AI adoption that is additive rather than disruptive.

LangSmith's agent improvement loop is a systematic process designed to enhance AI agents by utilizing traces as the foundational element for iteration and development. The loop involves collecting traces from various sources like production, staging, and test runs, and then enriching these traces with evaluations and human feedback to identify and address failure patterns. This process includes making targeted changes, validating improvements through offline evaluations, and deploying updates, which are continuously monitored by online evaluators to prevent regressions. LangSmith facilitates this cycle by providing tools for automated data generation and human annotations, ensuring that both qualitative and quantitative aspects of agent performance are addressed. The ultimate goal is to create a consistent feedback loop that iteratively improves agent reliability and functionality by leveraging enriched trace data, making it a central aspect of AI development and optimization within the LangSmith platform.

Victor Moreira, a Deployed Engineer at LangChain, presents a comprehensive checklist for evaluating AI agents, emphasizing the importance of agent evaluation, which differs from traditional software testing. The guide outlines a systematic approach to building, running, and optimizing agent evaluations by starting with simple end-to-end evaluations to establish a baseline and gradually adding complexity based on evidence of failure. Key components include defining clear success criteria, separating capability evaluations from regression evaluations, identifying failure causes, and ensuring evaluation ownership by a domain expert. The process involves using tools like LangSmith for trace analysis, categorizing failures, and designing specialized graders for different evaluation dimensions. The article highlights the significance of offline, online, and ad-hoc evaluations, promoting successful evaluations into regression suites, and integrating them into CI/CD pipelines to maintain agent reliability. It stresses the need to iterate continuously by adapting evaluations based on production feedback and evolving test suites when pass rates plateau.

Agent harnesses are essential components for connecting language models (LLMs) to their environments, enabling them to perform tasks effectively. By using agent middleware, developers can build on the foundations provided by LangChain and Deep Agent to create customized agent harnesses tailored to specific applications. Middleware in this context refers to a system that allows for the insertion of custom logic at various stages of an agent's operation, such as before or after model calls, enabling functionalities like context management, dynamic tool selection, and production readiness. LangChain offers prebuilt middleware options for common tasks like summarization and PII detection, and also allows developers to create their own middleware for specialized needs. The flexibility of middleware supports the decoupling of business logic from core agent code, facilitating the reuse of logic within organizations, and ensuring that agents remain adaptable to evolving requirements.

Kensho, S&P Global's AI innovation hub, addresses the challenge of efficiently navigating the company's vast and structured data estate by developing Grounding, a multi-agent framework that unifies data retrieval across various business units. Grounding serves as a centralized access layer for AI applications, ensuring insights are derived from verified datasets and enabling natural language queries against S&P Global's financial data. This system simplifies data access for financial professionals, eliminating the need for navigating complex schemas or specialized query languages, and provides real-time, citation-backed responses. The architecture leverages LangGraph to intelligently route queries to specialized Data Retrieval Agents (DRAs) across domains like equity research and macroeconomics, facilitating coherent insights by aggregating distributed responses. Kensho's custom DRA protocol ensures consistent data access patterns and accelerates collaboration across its multi-agent ecosystem, allowing rapid deployment of specialized financial AI products. The development process highlighted best practices in observability, multi-stage evaluation, and protocol optimization, which enhance system efficiency and maintain the reliability required in financial services.

Deep Agents, an open-source and model-agnostic agent harness, focuses on improving agent behavior by curating targeted evaluations, or "evals," that directly measure the desired behaviors of agents in production environments. By sourcing data from dogfooding, external benchmarks, and custom-written tests, Deep Agents ensures that each eval is designed to reflect real-world tasks and is self-documented with detailed explanations and categorized tags for efficient grouping and analysis. The approach emphasizes quality over quantity, cautioning against an excessive number of evals that might not accurately represent agent capabilities in production. The evals are run using pytest with GitHub Actions, focusing on correctness and efficiency metrics such as step ratio, tool call ratio, and solve rate, which help in refining model harnesses and optimizing agent performance. This methodology not only enhances agent reliability but also helps in efficiently managing resources by concentrating on the aspects that truly impact user experience and cost-effectiveness, ultimately fostering a shared responsibility among team members for maintaining and improving evals.

Fleet now offers shareable skills, allowing agents within a team to be equipped with specialized knowledge for specific tasks, which can be created from prompts, templates, or previous chats, and shared across workspaces to stay in sync. These skills function as persistent briefing documents, helping agents apply domain knowledge efficiently and consistently, ensuring that expertise remains within the team even as members change. Fleet provides multiple methods to create skills, including AI-assisted generation, manual writing, and templates for common tasks, enabling seamless onboarding and continuity. Skills are portable and can be integrated into development environments using the LangSmith CLI, ensuring consistent knowledge deployment across different platforms. Upcoming enhancements include version control and multi-owner permissions, fostering collaborative maintenance and precise task execution as agents engage in more complex, high-stakes work.

Moda is an AI-native design platform aimed at non-designers, providing tools to create professional-grade presentations and visual content without requiring a design background. It distinguishes itself from similar platforms like Canva and Figma by integrating a Cursor-style AI sidebar and leveraging a multi-agent system built on Deep Agents, with LangSmith offering observability. The platform addresses the challenge of AI in visual design by developing a custom context representation layer, which simplifies how layout information is processed, thus improving output quality and reducing token costs. Moda's AI system comprises three agents: the Design Agent for real-time design creation, the Research Agent for content gathering, and the Brand Kit Agent for maintaining brand consistency. These agents operate using a shared architecture that includes a triage step, a main agent loop, dynamic context loading, and full tracing in LangSmith. Moda's approach allows users to interact with a fully editable 2D vector canvas, transforming the user-AI dynamic into a collaborative process rather than a simple generate-and-replace model. The platform has found early success with B2B companies needing quickly produced, brand-aligned sales materials, and future developments include enhancing memory capabilities, completing migration to Deep Agents, and expanding brand context support for larger enterprises.

LangSmith Fleet was launched to facilitate the management, creation, and use of agents, introducing two types of agent authorization: Assistants and Claws. Assistants operate "on-behalf-of" users, utilizing their credentials to access data, whereas Claws possess their own fixed credentials, allowing broader usage across different channels like Slack, Gmail, and Teams. This dual authorization approach addresses various user needs, whether requiring personalized or shared agent interactions. Real-world examples include an Onboarding Agent acting as an Assistant with user-specific credentials, an Email Agent functioning as a Claw with fixed credentials, and a Product Agent monitoring competitors with its own Notion account. LangSmith Fleet also incorporates human-in-the-loop safeguards for sensitive actions and plans to enhance memory permissions for more granular control in future developments.

Google Cloud Next 2026 in Las Vegas will feature LangChain prominently at Booth #5006 in the Mandalay Bay Convention Center from April 22-24, offering demos and technical conversations about the latest in agent development. Presentations include updates on the LangChain ecosystem, showcasing tools like LangSmith Fleet, Studio, and agent observability capabilities. Attendees can engage with LangChain CEO Harrison Chase for discussions on AI application development challenges or meet the engineering team for insights on scaling multi-agent systems. A notable breakout session will cover building secure, scalable AI agent environments on Google Kubernetes Engine, including customer insights from Victor Moreira and Google Cloud experts. LangChain will also participate in a panel discussion about eliminating friction in developer experiences through AI agents and open standards. LangSmith is now available on the Google Cloud Marketplace, simplifying procurement for teams building AI applications. A happy hour co-hosted by LangChain, MongoDB, and Confluent offers networking opportunities for those interested in modern AI infrastructure.

LangSmith Fleet is an enterprise platform designed to facilitate the creation, management, and deployment of multiple agents within an organization, allowing teams to automate repetitive tasks efficiently. The platform's features include a robust agent identity and sharing model, enabling users to assign and control permissions, and manage how agents authenticate with external tools. Initially launched as LangSmith Agent Builder, the service has evolved to allow non-engineers to generate agents through simple prompts, expanding use cases from basic tasks to more complex operations. Key capabilities include centralized agent management through an "Agent Inbox" for oversight, agent observability with detailed tracing for audit purposes, and flexible authentication options that cater to different organizational needs. LangSmith Fleet's architecture supports seamless collaboration by offering tiered permissions and customizable sharing options, ensuring that agents are both secure and adaptable to various team requirements.

Polly is an AI assistant designed to aid in debugging complex agent traces within LangSmith by efficiently analyzing extensive steps and prompts to identify issues and provide insights. Now generally available, Polly has been integrated across all LangSmith pages, enhancing its capabilities and accessibility. This AI assistant not only retains context across navigation but also performs actions like updating prompts, creating datasets from failed runs, and writing evaluator code, effectively acting as a hands-on team member. By following workflows seamlessly, Polly helps users manage intricate debugging tasks, understand user sentiment in conversations, and refine evaluator logic. It also assists in making data-driven decisions by comparing experiment results and providing recommendations based on actual data. Developed through experience with teams building production agents, Polly addresses recurring challenges in the debugging process and supports users by handling tasks that typically slow down engineering efforts. Users can easily access Polly on LangSmith by setting up an API key, enhancing their ability to monitor and improve their workflows.

Over the past year, several engineering organizations, including Stripe, Ramp, and Coinbase, have developed internal coding agents to assist their development teams, integrating them into existing workflows through platforms like Slack, Linear, and GitHub. These agents share common architectural patterns, such as isolated cloud sandboxes, curated toolsets, subagent orchestration, and rich context integration, which ensure effective deployment in production environments. Open SWE, an open-source framework, captures these patterns and provides a customizable foundation built on Deep Agents and LangGraph, offering core architectural components seen across these implementations. The framework supports isolated execution environments, a curated toolset, context engineering, and subagent orchestration, allowing organizations to adapt the system to their specific needs. Open SWE serves as a starting point for teams exploring internal coding agents, providing flexibility through pluggable components, middleware for reliable orchestration, and options to customize sandbox providers, models, tools, triggers, and validation processes, while maintaining an upgrade path through continuous improvements to Deep Agents.

LangSmith Sandboxes, now in Private Preview, offer secure and scalable environments for executing untrusted code, addressing the risks associated with running arbitrary code without isolation. These sandboxes provide ephemeral, locked-down settings where agents can safely execute code while controlling resource access and consumption. With the LangSmith SDK, users can easily create a sandbox environment by integrating their API key, facilitating projects like Open SWE. Unlike traditional containers meant for vetted applications, LangSmith Sandboxes cater to unpredictable agent-generated code, aiding in tasks such as coding assistance, CI-style operations, and data analysis. Sandboxes integrate seamlessly with the LangSmith Platform, supporting persistent state, real-time output streaming, and features like pooling and autoscaling. Offering security enhancements like microVM isolation and an authentication proxy, LangSmith Sandboxes ensure sensitive information remains protected. Future developments include shared volumes, binary authorization, and comprehensive execution tracing to enhance security and auditability.

The deploy CLI, part of the langgraph-cli package, introduces a streamlined process for deploying and managing agents from the command line, with the initial command, langgraph deploy, enabling single-step deployment to LangSmith Deployment. This integration facilitates the incorporation of LangSmith Deployment into CI/CD workflows using platforms like GitHub Actions, GitLab CI, or Bitbucket Pipelines, by automating the creation of a Docker image and infrastructure setup, including Postgres and Redis services. Additional commands allow users to manage deployments, such as listing, viewing logs, and deleting deployments, while new templates for deep and simple agents can be generated with langgraph new. The latest version of langgraph-cli, featuring these new commands, is available for use, and users are encouraged to provide feedback to enhance the developer experience.

LangChain, an agent engineering company known for its successful open-source frameworks, has partnered with NVIDIA to create an advanced AI development platform designed for enterprises to build, deploy, and monitor AI agents at scale. This collaboration combines LangChain's LangSmith platform and various open-source tools with NVIDIA's Agent Toolkit, including models and microservices, to streamline the development process and enhance the capabilities of AI agents. The platform aims to reduce the development time and infrastructure complexity by offering comprehensive tools for agent orchestration, task planning, and resource management, alongside NVIDIA's optimized execution strategies that ensure high throughput and reduced latency. Furthermore, LangChain's integration into NVIDIA's Nemotron Coalition highlights a commitment to advancing open AI models and responsible AI practices, emphasizing content safety and policy compliance. This partnership is expected to support enterprises in moving rapidly from prototypes to full-scale production, offering an open and flexible solution tailored to diverse workflows.

Deep Agents has introduced a new feature in its SDK and CLI that allows AI models to autonomously compress their context windows, thus managing their working memory more efficiently. This tool enables models to replace older messages with a condensed summary of relevant information, optimizing context window usage without user intervention. Traditionally, context compression was manually triggered at fixed thresholds, which could be suboptimal during complex tasks. However, the new feature empowers models to determine the optimal times for compaction, enhancing workflow efficiency and reducing the need for manual tuning. The tool retains recent messages while summarizing preceding ones and is currently enabled in the CLI and available as an opt-in middleware in the SDK. Testing has shown that models are conservative about using the feature, typically choosing moments that enhance task performance. This development reflects a shift towards giving AI models more autonomy over their memory management, reducing reliance on rigid, predefined rules.

Harness engineering involves creating systems around models to make them function effectively as agents by incorporating code, configuration, and execution logic that the model itself lacks. Essential components of a harness include system prompts, bundled infrastructure, orchestration logic, and middleware, which help manage aspects such as durable state, code execution, and environment setup. By integrating features like filesystems, bash tools, sandboxes, and memory management, harnesses enable models to perform tasks they cannot achieve independently, such as maintaining state across interactions and accessing real-time knowledge. Harnesses support continuous learning, self-verification, and long-horizon execution, enhancing a model’s capability to autonomously solve complex problems. As models become more advanced, some harness functions might be absorbed into the model itself, yet harness engineering remains crucial for optimizing model performance by providing well-configured environments and tools. This domain continues to evolve, exploring areas such as dynamic tool assembly and parallel agent orchestration, aiming to refine the interaction between model intelligence and system design for more effective agent development.

In the evolving landscape of software development, the roles within Engineering, Product, and Design (EPD) are undergoing significant transformation due to the rise of coding agents that simplify the code writing process. The traditional reliance on Product Requirement Documents (PRDs) is diminishing as coding agents allow for rapid prototyping, shifting the bottleneck from implementation to review. Consequently, the focus is now on ensuring that code is well-architected, user-centric, and intuitively designed, with EPD professionals acting as critical reviewers rather than just implementers. Generalists, who can navigate product, engineering, and design, are increasingly valuable, as they can expedite processes by minimizing communication overhead. Meanwhile, the necessity for system thinking and product sense has intensified, requiring EPD roles to blend and adapt, with specialists needing to excel in fast-paced reviews and communication. The increased accessibility of coding has democratized building, allowing various roles to see themselves as equally empowered by these tools, while the archetypal EPD professional now sits at the intersection of culture and technology, possessing an intuitive grasp of product potential and technical possibilities.

LangChain developed a GTM agent to streamline the sales process by automating lead research and email drafting, significantly increasing efficiency and conversion rates. The agent integrates with various platforms, such as Salesforce and Gong, to gather and process data, ensuring that sales representatives have the necessary context and insight before reaching out to leads. Built on Deep Agents, it handles long-running, multi-step processes by orchestrating multiple tools and data sources, offering features like relationship-aware personalization, explainability, and a learning loop from representative edits. The system logs every action to evaluate quality and catch regressions, expanding its capabilities to include account intelligence, which identifies deal risks and opportunities. Initially for sales, the agent's broad data access led to its adoption across other departments, demonstrating its versatility and effectiveness. This adoption was facilitated by the agent's ability to perform complex orchestration beyond simple LLM calls, leveraging infrastructure designed for robust workflows. The GTM agent's development emphasizes starting with clear success criteria, ensuring human-in-the-loop processes, and integrating deeply with existing systems to foster organic usage expansion.

LangChain has been developing skills to enhance the performance of coding agents like Codex, Claude Code, and Deep Agents CLI within its ecosystem, focusing on how these skills can be effectively evaluated. Skills are dynamic, task-relevant instructions or scripts that aim to improve agent performance in specialized domains. They are crucial for optimizing an agent's capabilities without overwhelming it with unnecessary tools, which could degrade performance. The evaluation process involves defining specific tasks, employing skills to aid in their completion, and then assessing performance improvements. A clean testing environment is essential to ensure consistent and reproducible results, with metrics such as task accomplishment rate, skill invocation, and task completion speed tracked using LangSmith evaluations. The content of skills should be modular and strategically placed to ensure reliable invocation, with AGENTS.md and CLAUDE.md files providing consistent guidance. Testing different skill configurations revealed that while skills generally enhance task completion rates, understanding why agents fail is vital for iterative improvements. Integration with LangSmith provides observability into the agents' actions, facilitating faster iteration and refinement of skills.

A new CLI is being introduced alongside a set of skills to enhance AI coding agents' capabilities within the LangSmith ecosystem, significantly improving performance metrics from 17% to 92% on certain tasks. The LangSmith CLI enables coding agents to efficiently interact with LangSmith by providing tools for tracing, dataset curation, and experiment execution, thus facilitating a terminal-first approach that is deemed crucial for future agent development. Skills, which are dynamically loaded based on task relevance, help improve agent performance without overwhelming them with unnecessary tools and are both portable and shareable. The initial set of LangSmith skills focuses on tracing, dataset building, and evaluation, aiming to create a virtuous cycle in agent development by allowing agents to add tracing logic, debug behavior, build testing datasets, and validate correctness through evaluations. Installation options are available for both local and global projects, and the initiative encourages community feedback and contributions for future skill enhancements.

AI coding agents are being enhanced through the introduction of skills specifically for the open-source LangChain ecosystem, significantly improving performance from 29% to 95% in tasks related to LangChain, LangGraph, and DeepAgents. These skills, which are curated instructions and resources, are designed to be dynamically loaded only when relevant, addressing the historical issue of performance degradation when too many tools are available to an agent. The skills are portable, shareable, and are stored as markdown files and scripts that can be accessed on demand. They are divided into three categories: LangChain, LangGraph, and DeepAgents, and are available for integration into any coding agent that supports this functionality. Installation is simplified through the use of commands such as npx skills, which allows for both local and global integration. The introduction of these skills is part of a broader initiative to enhance the LangChain and LangSmith ecosystems, with plans to expand skills content as new capabilities emerge. The community is encouraged to contribute ideas for additional skills and improvements.

February saw a range of updates from the LangChain team, including enhancements to the Agent Builder, such as a central chat for all tools and a unified tool registry, and the ability to schedule reports with the Insights Agent. Additionally, new features allow users to configure trace inputs and outputs and pin experiments as baselines for performance comparison. On the open-source front, `deepagents` v0.4 introduced sandboxed environments for running agents. The team also announced the LangChain Academy course on building reliable agents and shared customer success stories, such as monday.com accelerating their evaluation feedback loop by using LangSmith. Upcoming events include various community meetups and tech talks in cities like San Francisco, London, and New York, focusing on topics like agent production monitoring, harness engineering, and the future of AI agents.

Clay is a platform designed to enhance the efficiency of go-to-market teams by enabling them to build, enrich, and activate lists of companies and people through AI-powered tools. It serves a diverse clientele, from startups to large enterprises, facilitating tasks such as sourcing target accounts, qualifying leads, and drafting personalized outreach. With approximately 300 million AI agent runs monthly, Clay's operations involve complex multi-step processes like web scraping and data synthesis. To manage scalability challenges like quality assurance, cost control, and the rapid pace of model releases, Clay relies on LangSmith for observability and evaluation. This integration allows for real-time monitoring, debugging, and structured evaluations, helping Clay maintain cost reconciliation and improve its pricing strategy. LangSmith's capabilities provide critical insights into usage patterns, error rates, and model performance, enhancing Clay's ability to adapt quickly to new AI model introductions and maintain operational efficiency.