April 2026 Summaries

Madrigal Pharmaceuticals has developed a flexible and scalable multi-agent research and intelligence platform using LangChain and LangSmith to integrate, search, and synthesize information from diverse datasets at scale, particularly for pharmaceutical applications. By abstracting data sources and employing modular skills, Madrigal transformed single use cases into a comprehensive platform, significantly reducing development time for new applications and maintaining deployment simplicity. The platform ensures observability and learning from real-world errors through LangSmith's tracing and evaluation capabilities, enabling the system to adapt and improve continuously. LangSmith's managed deployment, tracing, and evaluation features allow Madrigal to quickly transition from prototypes to enterprise use, enhancing productivity and impact, particularly in the fight against metabolic dysfunction-associated steatohepatitis (MASH). The orchestrator's role in coordinating multiple agents, each with specific tasks, allows for parallel processing, increasing efficiency and accuracy while maintaining flexibility. LangChain and LangSmith's framework supports rapid scaling across domain experts, enabling Madrigal to leverage their expertise and improve the platform iteratively.

Deep Agents has introduced model-specific profiles to optimize prompts, tools, and middleware for different model families, such as OpenAI, Anthropic, and Google models, enhancing performance by 10-20 points on the tau2-bench. Previously, Deep Agents operated with a generic set of parameters that worked across all Large Language Models, but the new harness profiles allow for tailored adjustments, improving efficiency and adherence to model-specific prompting guides, such as OpenAI's Codex and Anthropic's Claude. This customization is vital because different models respond differently to specific prompting techniques, as demonstrated by improved results on Terminal-Bench 2.0, where adjustments elevated gpt-5.2-codex from 52.8% to 66.5%. The profiles are declared as override layers for system prompts, tool inclusion, and naming, among other factors, and can be registered or overridden by users to suit specific needs, while maintaining consistent call sites. The introduction of these profiles empowers builders to manage and test agent configurations effectively, with the goal of creating optimal harnesses for various tasks.

In anticipation of the upcoming Interrupt conference, Jake Broekhuizen, an MC with a background in electrical engineering and computer science, is highlighted for his expertise in agent engineering and his role at LangChain, a company focused on integrating AI into enterprise workflows. Jake's journey from ServiceNow, where he first tackled the challenge of connecting data to large language models like OpenAI's, to his current position at LangChain, showcases his commitment to enhancing AI deployment strategies. The conference, now in its second year, aims to delve deeper into the evolution of agent engineering by exploring how enterprises can effectively scale agents and integrate them into business operations. With a focus on real-world applications and industry trends, the event promises exciting announcements in AI agent development, including advancements in observability and self-improving systems. The broader landscape sees a shift towards specialized agent harnesses tailored to specific domains and the management of agents on a fleet scale, reflecting a pivotal change in how businesses leverage AI technologies.

With the EU AI Act compliance deadline approaching on August 2, 2026, LangSmith and LangChain OSS offer solutions to help developers meet the stringent requirements for high-risk AI systems, which include fields like financial services, healthcare, and critical infrastructure. The Act mandates a comprehensive risk management system, automatic event logging, transparency, human oversight, and ongoing monitoring for AI systems that perform complex decision-making processes. LangSmith provides tools for full observability and evaluation, capturing every step of an agent's execution with detailed trace data to ensure compliance with articles covering risk management, data governance, and human oversight. The platform supports continuous quality scoring, customizable evaluators to measure bias, toxicity, and accuracy, and human-in-the-loop features to allow intervention and oversight. By offering deployment options that adhere to EU data residency requirements, LangSmith ensures that AI systems can be managed and audited effectively, aligning with regulatory expectations while providing operational control.

LangChain's April 2026 newsletter highlights significant updates and upcoming events, emphasizing advancements in their agent development platform, LangSmith, which now includes over 30 evaluator templates and cost alerting features to monitor expenses. The newsletter introduces Deep Agents, an open-source, model-agnostic system that allows users to maintain ownership of their data, and discusses the deployment capabilities of "deepagents deploy" for scalable server setups. The upcoming Interrupt 2026 event in San Francisco is set to feature industry leaders discussing the future of AI agents, with presentations from teams at companies like Cisco and LinkedIn. LangChain also shares success stories, such as Credit Genie's AI assistant developed with LangGraph and LangSmith, which identified a gap in customer support, highlighting the platform's value in optimizing agent performance. Additionally, the newsletter promotes various meetups and webinars across North America to engage with the community and share insights on deploying deep agents effectively.

Credit Genie, a mobile-first financial wellness platform, integrated LangSmith's Insights Agent into their AI financial assistant, AskGenie, to enhance user interaction analysis and identify gaps in service. By running two configurations of the Insights Agent, one with predefined categories and another allowing for automatic category generation, Credit Genie discovered unexpected patterns such as a high volume of customer support requests that their existing chatbot did not address. These insights led to the development of new features including cash advance status lookups and repayment date modifications, and the integration of in-chat support requests. The continuous use of Insights Agent has become integral to Credit Genie's product development process, enabling the team to quickly adapt to behavioral trends, enhance the AskGenie experience, and maintain rapid iteration cycles in response to user needs.

The conceptual guide outlines the infrastructure and runtime requirements for deploying long-running, production-grade deep agents, emphasizing the need for durable execution, memory management, multi-tenancy, human-in-the-loop (HITL) capabilities, and observability. It introduces the LangSmith Deployment (LSD) and Agent Server as the core runtime components, which handle execution, memory, concurrent message handling, and provide open, model-agnostic protocols like MCP and A2A for integration. Durable execution ensures agents can pause and resume tasks, handling crashes and human interruptions seamlessly, while memory is categorized into short-term (conversation-specific) and long-term (cross-conversation) storage. Multi-tenancy is addressed through robust authentication and authorization mechanisms, while HITL supports dynamic interruption and resumption of tasks for human oversight. Real-time interaction challenges like streaming and concurrent messaging are tackled with strategies such as enqueueing and interrupting ongoing processes. Observability is crucial for understanding agent behavior, with tracing and time travel features enabling detailed analysis and debugging. Code execution is facilitated through sandboxed environments to maintain security, and integration capabilities via open protocols ensure seamless connectivity with existing systems. The guide highlights the open-source nature of deepagents deploy, allowing for customization and avoiding vendor lock-in, and it underscores the importance of an iterative development cycle to continuously improve agent performance.

Agentic engineering is a transformative approach in software development, where AI agents function as coordinated digital team members with specific roles, shared memory, and a unified observability system, aiming to expedite software delivery beyond mere code generation. A pilot study demonstrated a 93% reduction in time-to-root-cause in debugging workflows, saving over 200 engineering hours, while development workflows saw a 65% reduction in execution time, primarily by streamlining downstream testing rather than speeding up code generation. Unlike AI coding agents such as Codex that focus on translating intent into code within isolated sessions, agentic engineering operates at a higher abstraction level, orchestrating cross-team workflows with long-term memory and state management capabilities. This system, leveraging LangChain's tools like LangGraph and LangMem, mirrors real-world engineering teams, enabling a structural shift in software delivery by minimizing coordination overhead, reducing cross-team latency, and optimizing human attention, thereby enhancing resilience and scalability across the software lifecycle.

LangSmith has introduced reusable evaluators and evaluator templates aimed at enhancing agent evaluation processes, which are critical for debugging and improving agent performance. The platform now offers over 30 templates that focus on safety, response quality, trajectory, user behavior, and multimodal evaluations, allowing users to either use them as-is or customize them for specific needs. These templates facilitate both online monitoring and offline experiments, helping categorize production traffic and assess experiments. A centralized Evaluators tab allows users to manage and apply evaluators across multiple tracing projects, ensuring consistency and eliminating the need for duplicate evaluators. This update is part of LangSmith's broader effort to streamline evaluation by allowing users to build evaluators once and deploy them universally, while upcoming features will include spend visibility to help track evaluation costs.

Siddhant Dash, a Senior Product Manager at Cisco AI Defense, discusses the importance of securing LangChain agents using middleware as the enforcement point for agent security. Middleware allows for a clean integration that keeps LangChain code uncluttered while providing a consistent point for applying security policies across the agent loop. Cisco AI Defense offers two modes: monitor, which records risk signals and decision traces without interruption, and enforce, which blocks policy violations with an auditable reason. The protection spans across LLM calls, MCP tool calls, and middleware, essential for multi-agent systems where orchestrators link agents at runtime. The article emphasizes the necessity of clear enforcement points to apply policies and keep an auditable record, particularly as LangChain facilitates quick transitions from prototypes to functional agents capable of interacting with sensitive systems and data. Cisco AI Defense's integration into LangChain through middleware provides a consistent runtime contract, and the organization is contributing this integration upstream via LangChain’s middleware framework, inviting feedback and collaboration from users.

Async subagents have been introduced to address limitations in traditional agent architecture, particularly when managing complex and lengthy tasks. Unlike inline subagents that block the supervisor agent by requiring synchronous tool calls, async subagents allow for non-blocking operations by returning a task ID immediately, enabling supervisors to manage multiple tasks simultaneously. This approach, based on the Agent Protocol, allows the supervisor to maintain control over the process by sending follow-up instructions, canceling unnecessary tasks, and providing updates mid-task, enhancing flexibility and efficiency. Async subagents operate independently with their own processes and state, offering deployment flexibility across different platforms, whether hosted on LangSmith deployments or self-hosted infrastructure. This advancement signifies a shift from a "fire-and-forget" to a "fire-and-steer" model, allowing agents to manage more complex workflows effectively.

Stale code samples in documentation are a universal issue, especially as APIs and dependencies evolve rapidly, but automating the testing of these samples can mitigate the problem. By using the Deep Agents CLI, companies like LangChain streamline the process of making inline code testable, significantly reducing manual work. This approach involves extracting code into standalone files, adding setup and teardown scripts, and employing a CI system to regularly run and test the code snippets. Deep Agents use "skills"—reusable instructions that automate tasks like migrating inline code to standalone, testable examples, ensuring documentation remains accurate and functional. This system not only automates the laborious process but also integrates testing seamlessly into documentation workflows, improving reliability and maintainability.

Agent harnesses have become the preferred method for building agents due to their integral role in managing agent memory, which is crucial for creating personalized and effective experiences. These harnesses are not just temporary scaffolding but essential systems that facilitate interactions between large language models (LLMs) and various data sources. The text discusses the importance of open harnesses that allow developers to own and control their agent's memory, as reliance on closed, proprietary harnesses can lead to a loss of control and lock-in to specific platforms. The article highlights the industry's ongoing evolution regarding memory management, noting that while the concept is still in its infancy, the ability to retain memory will become increasingly important as it enables the development of proprietary datasets and personalized experiences. To mitigate the risk of lock-in, open-source solutions like Deep Agents are being developed, allowing for model agnosticism and the use of open standards in memory storage, empowering developers to maintain ownership over their agent's memory and ensuring flexibility in choosing the most suitable models.

Interrupt 2026 is set to take place on May 13–14 at The Midway in San Francisco, building on the success of the previous year's conference which focused on the viability of deploying agents in production. This year, the event aims to address how to implement agents at an enterprise scale, exploring the necessary team structures, tooling, and infrastructure. Notable speakers include Harrison Chase, Co-founder and CEO of LangChain, Andrew Ng of DeepLearning.AI, and Chirantan “CJ” Desai, CEO of MongoDB, who will delve into advancements and future trends in agent technology. The conference will feature keynote speeches, fireside chats, and presentations from industry leaders like Lyft, Apple, and LinkedIn, each sharing insights into their agent platforms and evaluation systems. Attendees will have the opportunity to engage in workshops, participate in AMAs, and explore demo stations, all designed to facilitate learning and networking in the field of AI-driven enterprise solutions.

Rahul Verma, a Deployed Engineer at LangChain, emphasizes the importance of incorporating both documented and tacit human knowledge into AI agents to improve their performance and reliability. Using a financial services firm's "Copilot for traders" as a real-life example, the text illustrates how AI agents can automate workflows, like generating SQL queries, to free up data scientists and provide traders with quicker responses. To ensure these AI systems work effectively, they must integrate both financial domain knowledge and technical database insights, requiring input from domain experts. The text outlines a comprehensive approach to designing AI agents, including using deterministic code for critical steps, configuring tools with the right parameters, and employing context engineering for better information retrieval. It highlights the significance of incorporating human judgment into an iterative improvement loop involving development, monitoring, and testing, using automated evaluations aligned with expert judgment to efficiently refine agent performance. The LangSmith platform is mentioned as a tool to facilitate this process by providing features like Align Evaluator, annotation queues, and Insights Agent to gather real-time data and insights, ultimately creating a continuous cycle of improvement that leverages human expertise and automated evaluations to enhance AI agent functionality.

Deep Agents deploy is a newly launched beta platform designed to facilitate the deployment of open-source, model-agnostic agent harnesses in a production-ready manner, prioritizing user ownership of memory and customization. Built on Deep Agents, it allows for the creation of customized agents by integrating orchestration logic, tools, and skills, supporting a wide range of models and sandbox providers without locking users into a specific ecosystem. Key features include the ability to deploy agents with custom instructions and skills, operate in scalable multi-tenant environments, and integrate with open standards such as MCP, A2A, and Agent Protocol, while allowing self-hosting to ensure data and memory remain under user control. This approach contrasts with other offerings like Claude Managed Agents, which can create substantial lock-in by tying memory and context to closed APIs, leading to potential challenges in migrating models or harnesses. Deep Agents deploy aims to provide flexibility and control over agent memory and model selection, emphasizing an open ecosystem that avoids the pitfalls of proprietary lock-in.

Better-Harness is a system designed to improve AI agents through a process of iteratively refining harnesses using evaluations (evals) as a learning signal, similar to training data in machine learning. The approach emphasizes the importance of high-quality evals, sourced from hand-curated examples, production traces, and external datasets, to guide agents towards desired behaviors and prevent overfitting. The system employs a cycle of data sourcing, experiment design, optimization, and review, with evals categorized by behavioral tags to enable targeted experiments and holdout sets to ensure generalization. By integrating human review and trace analysis, Better-Harness aims to enhance agent performance by discovering and addressing failure modes while maintaining a focus on generalization and avoiding regressions. The results from testing this system with models like Claude Sonnet 4.6 and Z.ai’s GLM-5 show improved agent behavior, demonstrating the potential for this approach to autonomously refine agent harnesses and adapt to various domains.

Deep Agents and deepagentsjs have introduced new minor versions that feature asynchronous (async) subagents and expanded multi-modal filesystem support. Async subagents enable tasks to run independently on remote servers without blocking the main agent's execution, allowing for parallel task management and stateful interactions across interactions. This enhancement addresses the limitations of inline subagents by enabling the supervisor to manage multiple tasks simultaneously and interact with the user without delays. The adoption of LangChain's Agent Protocol ensures seamless async task management and interoperability. Additionally, the release extends multi-modal support to include various file types such as PDFs, audio, and video, enhancing the agent's ability to process diverse data inputs without API changes. The updates aim to improve efficiency and flexibility in handling complex tasks and diverse data formats, making it easier for users to deploy and manage Deep Agents across different environments.

A new partnership between LangSmith Fleet and Arcade.dev is introduced, enabling LangSmith Fleet agents to access Arcade's extensive library of over 7,500 agent-optimized tools through a centralized and secure MCP gateway. This integration simplifies the process for teams to create, use, and share agents that autonomously operate across various tools like Salesforce, Notion, and Slack, providing a single access point for secure, efficient tool connectivity. The Arcade MCP gateway minimizes the integration burden by allowing users to connect with their own credentials, maintaining consistent tool structures and enhanced tool descriptions tailored specifically for language model selection and invocation. This collaboration supports diverse use cases by offering over 60 pre-built templates, handling secure tool authentication and authorization, and ensuring that actions align with user-specific permissions, thereby streamlining agent operations in environments with varying access levels.

Continual learning in AI systems involves not only updating model weights but also optimizing the harness and context layers. The model layer focuses on updating model weights, often facing challenges like catastrophic forgetting. The harness layer refers to the code and tools driving the agent, which can be optimized through methods like Meta-Harness, where logs are evaluated, and improvements are suggested. The context layer, or memory, includes instructions and skills that can be updated at various levels, such as agent, user, or organization, to enhance configuration and performance. These updates can occur offline or in real-time, with traces of the agent's execution path playing a crucial role in driving improvements across all three layers. Platforms like LangSmith help collect and use these traces to refine models, harnesses, and context, ensuring that systems like Deep Agents can support ongoing learning and memory updates effectively.

Vishnu Suresh, a software engineer at LangChain, describes the development of a self-healing deployment pipeline for the GTM Agent that automates regression detection, triage, and fixes through the use of an internal coding agent, Open SWE. The system leverages GitHub Actions to capture build and server logs, with automated processes identifying and addressing issues without manual intervention until review. The pipeline distinguishes between build failures, which are straightforward to detect, and more complex server-side errors, which require statistical analysis and triage to differentiate genuine regressions from background noise. By using a Poisson test to model expected error rates and a triage agent to establish causality, the system effectively closes the loop from error detection to resolution. Future improvements being considered include widening the lookback window for error attribution, enhancing error grouping methods using vector space clustering, and balancing between fixing forward and rolling back based on severity and confidence. The self-healing approach is expected to become increasingly common, allowing for faster deployments and reducing the need for constant manual monitoring.

Open models like GLM-5 and MiniMax M2.7 are proving to be effective alternatives to closed frontier models in performing core agent tasks such as file operations, tool use, and instruction following, offering similar performance at significantly reduced costs and latency. Evaluations demonstrate that these open models can effectively be deployed in production environments, providing consistent and predictable performance that supports real-world workflows. Despite being smaller than their closed counterparts, open models benefit from specialized infrastructure that optimizes latency and throughput, making them a cost-effective choice for latency-sensitive applications. Evaluation methods assess correctness, solve rate, step ratio, and tool call ratio, highlighting that open models achieve competitive results with efficiency. Additionally, Deep Agents supports easy integration and comparison of open models, simplifying the process of adapting these models into existing systems and allowing for dynamic model switching to balance planning and execution tasks.

LangChain's latest updates include a new NVIDIA integration, the rebranding of Agent Builder to LangSmith Fleet, and the introduction of LangSmith Sandboxes in Private Preview, which provide secure environments for agents to run code. LangSmith Fleet now features agent identity, sharing, and permissions, as well as skills that equip agents with specialized task knowledge. The new LangGraph Deploy CLI enables seamless deployment of agents from the terminal, while Attribute-Based Access Control (ABAC) enhances resource access management. The company also released langgraph v1.1 and deepagents v0.5.0 alpha, introducing new features such as type-safe streaming and multi-modal support. LangChain is hosting workshops and events, including Interrupt 2026 in San Francisco, where industry leaders like Jensen Huang and Andrew Ng will share insights. Additionally, LangChain's collaboration with NVIDIA aims to offer a comprehensive platform for building and running production agents. Community events are taking place globally, fostering discussions around LangChain's technology stack and applications.