May 2026 Summaries

Interpreter skills are an innovative extension to agent skills, allowing agents to include a TypeScript module within a skill to execute code directly, enhancing their capability to perform complex tasks more efficiently and predictably. This development, introduced to Deep Agents, leverages a TypeScript runtime for agents to express multi-step work as code, enabling them to take on more deterministic routines while maintaining discretion on when to apply these routines. By importing skill modules and executing them, agents can perform tasks such as repository triage with fixed procedures, thus ensuring reliability and accuracy. Interpreter skills blend the flexibility of modern agent harnesses with the deterministic nature of traditional workflows, enabling agents to handle intermediate states and complex operations effectively. This approach allows for the creation of testable, versioned, and reusable libraries of best practices, providing a more structured and reliable method for agents to perform tasks while keeping the model in charge of strategic decision-making.

LangSmith has introduced significant updates at Interrupt 2026, including the launch of the LangSmith Engine and the general availability of LangSmith Sandboxes, both designed to enhance agent development and execution. The LangSmith Engine optimizes the agent development lifecycle by monitoring production traces, identifying recurring failures, and suggesting fixes, while Sandboxes offer secure environments for agent code execution. Additional updates focused on boosting agent infrastructure, observability, and governance. The event showcased multiple talks from industry leaders, all now accessible on demand, and introduced LangChain Labs, a new initiative aimed at advancing continual learning for agents. Upcoming events are scheduled in various locations, offering technical discussions and workshops to further explore agent development and integration.

Lyft's SCX Data Science and MLE team has innovated a self-serve AI platform that empowers non-technical domain experts to independently develop and refine AI agents, significantly expediting the agent development process from six months to just a few weeks. Utilizing LangGraph's multi-agent architecture and LangSmith's tracing and monitoring tools, the platform supports seamless integration for handling complex customer support issues across various categories like account access and damage claims. This system enables domain experts to create configurable agents using natural language prompts, while maintaining high-quality standards through an automated LLM-as-a-judge evaluation pipeline. The initiative has enhanced operational efficiency and increased AI resolution rates by 16%, highlighting the potential for expanded global deployment and continuous improvement in prompt quality and evaluation coverage.

Self-hosting LangSmith on Kubernetes offers platform teams enhanced control over infrastructure and security while managing complex deployments through various tools like Helm and kubectl, which can lead to frequent context switching. Mission Control, an in-cluster application, simplifies these operations by providing a centralized interface for deploying, configuring, observing, and troubleshooting LangSmith and LangChain infrastructure. It enhances efficiency through features such as a guided onboarding flow, a bidirectional Helm values editor, cluster-aware preflight checks, health and observability tools, and version-aware deployment management. Additionally, Mission Control includes an operator assistant for LangSmith-specific queries and provides rule-based alerts and unified search capabilities to streamline incident response and operational management. By integrating these functionalities within the Kubernetes framework, Mission Control minimizes manual intervention and context switching, offering a cohesive solution for teams operating in private or regulated environments.

In large enterprises, the deployment of corporate laptops is typically accompanied by a suite of security tools aimed at preventing unauthorized access and data leaks, given developers' extensive access privileges. This environment becomes more complex with the introduction of agents, which can replicate developer capabilities on a massive scale, necessitating even stricter security measures. LangSmith Sandboxes address this by creating isolated environments for agents, coupled with a sandbox auth proxy that manages interactions with external services without exposing sensitive credentials within the runtime. This setup ensures that agents have the necessary access to APIs and services while maintaining security by injecting credentials at the network layer, rather than embedding them within the runtime environment. This approach not only keeps credentials secure but also allows for granular control over network access, ensuring agents can only interact with authorized services. Additionally, the proxy's ability to handle dynamic credentials and enforce egress policies further enhances security by preventing unauthorized internet access and ensuring compliance with infrastructure policies. This model supports scalable agent deployment while minimizing the risks associated with credential exposure and network vulnerabilities.

Recent advancements in Deep Agents and related technologies like LangChain and LangGraph are revolutionizing how agent-based applications handle streaming data by focusing on structured application events rather than raw data chunks. These developments enable agents to efficiently plan, delegate tasks, use tools, and manage state while providing real-time, multimodal data streams that can be easily rendered across various frontend frameworks like React, Vue, Svelte, and Angular. The new streaming model is built around typed events, allowing developers to subscribe to specific channels and namespaces without downloading unnecessary data, thus optimizing performance and user experience. This approach facilitates rendering live updates of complex agent work, including reasoning, media, and tool calls, in a structured manner, which is essential for creating scalable, interactive user interfaces for long-running agent applications. Additionally, the release of framework SDKs provides tools for building these applications efficiently, offering features like scoped subscriptions, content blocks, and projections, which streamline the integration of custom domain-specific streams and enhance overall system reliability and usability.

Deep Agents introduces interpreters, small embedded runtimes that allow agents to write and execute code within the agent loop, offering a middle ground between one-at-a-time tool calls and full sandboxes. This functionality enables agents to express complex, multi-step workflows, maintain intermediate states outside of the model context, and execute actions more predictably, similar to a Python or Node REPL. Unlike traditional environments, interpreters start with a limited runtime, lacking APIs for filesystem, network, or shell access, and instead rely on explicit bridges for these capabilities. This design ensures a narrower action space and greater predictability, as additional functionalities must be deliberately added by the host. Interpreters facilitate efficient tool calling, manage control flow, and reduce token usage by retaining intermediate outputs within the runtime, rather than routing them back through the model. This setup is particularly beneficial for tasks involving large datasets, programmatic tool calling, and recursive orchestration, allowing agents to manage state more effectively and execute delegated tasks through a controlled API boundary. Deep Agents supports interpreters in both Python and TypeScript, with programmatic tool calling enabled through an allowlist, ensuring that only designated tools can interact with the interpreter.

Macro research desks require efficient identification and analysis of anomalous economic performance among countries, but the process is often hampered by fragmented data sources. To address this, an AI-driven agent was developed to streamline the analysis of 2025 GDP data for EU member states, identifying Ireland's significant growth and Germany's contraction, among other findings. The agent utilizes a methodical process involving multiple subagents, each handling distinct tasks like data retrieval, anomaly detection, and in-depth analysis. The architecture is built on the Finance Research API and LangSmith, ensuring traceability and transparency by logging every decision and data point back to its source. This system not only provides comprehensive reports but also maintains a detailed audit trail, crucial for compliance with regulations such as MiFID II, DORA, and the EU AI Act. The approach is designed to be modular and adaptable, supporting various financial research applications and allowing customization to fit specific workflows.

LangSmith Engine is an agent designed to enhance agent development by analyzing agent traces to identify recurring issues and suggest improvements. It operates as an orchestrator within the agent improvement loop, which includes building, testing, deploying, and monitoring agents. Engine identifies recurring failure patterns in traces, turns them into actionable issues, and proposes improvements such as evaluators and dataset examples. It uses a structured approach to handle large volumes of traces, employing a screening phase to identify suspicious traces and a deeper investigation phase for likely issues. Engine maintains an Issue Board where it records issues with evidence traces and proposes evaluators to catch similar patterns in future traces. The system relies on LangSmith CLI for interactions and uses a sandbox environment for operations, allowing it to adapt based on user actions and preferences. Through its architecture, Engine facilitates converting production failures into offline test coverage, enabling teams to improve their agents efficiently by focusing on recurring patterns and necessary fixes.

LangChain Labs has been launched as an applied research initiative focused on continual learning to advance open, applied research for agent technology. Collaborating with industry partners such as Harvey, Nvidia, and Baseten, LangChain Labs aims to improve the development and performance of agents by capturing and utilizing signals from agent runs. The initiative seeks to address challenges like extracting useful data from large-scale agent information, optimizing agent operations within organizational constraints, and simplifying the creation of evaluation and simulation environments. LangChain Labs emphasizes the importance of prompt optimization across model families, facilitating easier transitions and reducing manual tuning, in pursuit of a multi-model future. The initiative commits to sharing research, evaluations, and open-source integrations with the broader agent-building community to drive forward the development of efficient, self-improving agents.

Interrupt has launched a range of new products and features designed to streamline the agent development lifecycle, including the LangSmith Engine, SmithDB, Managed Deep Agents, LangSmith Sandboxes, Context Hub, and LangSmith LLM Gateway. The LangSmith Engine automates the process of diagnosing and fixing production issues, while SmithDB is a Rust-based database optimized for handling large volumes of agent trace data. Managed Deep Agents offer a hosted runtime environment for creating and managing complex agents, and LangSmith Sandboxes provide secure code execution environments. Context Hub centralizes the management of files that influence agent behavior, enabling collaborative updates without relying on GitHub. The LangSmith LLM Gateway enforces spending limits and safeguards sensitive data during interactions with LLM providers. Additionally, LangSmith Fleet has introduced sandbox capabilities for agents to execute more advanced tasks, supplemented by prebuilt templates for various agent types. These innovations collectively aim to enhance the efficiency, security, and collaborative potential of agent development and deployment.

LangSmith has introduced the LangSmith LLM Gateway in private beta, a governance layer designed to integrate seamlessly with the existing LangSmith framework where agents are built, observed, and evaluated. Positioned between agents and LLM providers, it enforces cost limits and redacts sensitive data, streamlining policy enforcement and observability without requiring external dashboards or tools. The Gateway offers features such as spend limits, real-time cost visibility, and sensitive data redaction, all within the LangSmith workspace to maintain trace continuity and ease policy management. This integration allows teams to detect, investigate, and remediate issues without leaving the LangSmith environment, reducing the need for multiple tools and context switches. By anchoring governance to the agent framework, the Gateway aims to provide a comprehensive control plane for runtime policies, covering not only LLM calls but potentially all aspects of agent interactions, with plans to expand security controls and enforcement flexibility.

LangSmith has introduced Context Hub, a centralized platform within its LangSmith suite aimed at managing, versioning, and collaborating on files critical to agent behavior, such as AGENTS.md files, skills, and policies. Context, which includes system instructions, examples, and domain-specific knowledge, plays a crucial role in shaping agent behavior, often more so than the agent's harness. Context Hub addresses the need for a distinct workflow since context is frequently managed by non-engineering teams and requires rapid updates. The platform offers features like versioning, tagging, and commenting to facilitate collaboration and ensure that the context remains a first-class component of agent systems. LangSmith's initiative, supported by partnerships with companies like Elastic, MongoDB, Pinecone, and Redis, aims to establish an open standard for agent memory, enhancing retrieval and context management workflows that are vital for deploying reliable, knowledgeable agents in production environments.

The latest DeepAgents release focuses on enhancing performance across several dimensions, including the model and agent layers, scalability, and longevity. Key features include a Code Interpreter that allows agents to execute tasks autonomously within a programmable workspace, improving workflow efficiency by keeping intermediate processes out of the model context. Harness profiles enable better tuning for open-weight models, enhancing model performance and cost-effectiveness. Streaming is introduced as a primitive for real-time application insights, allowing developers to track and manage agent activities with precision. DeltaChannel optimizes storage by saving only the differences at each checkpoint, significantly reducing the storage requirements for long-running agents. Additionally, the ContextHubBackend provides a versioned storage system for agent-related files, ensuring that improvements in context and skills are maintained across agent runs. These features collectively aim to optimize agent performance, reduce costs, and enhance the user experience by providing a more efficient and scalable framework.

LangSmith Engine is a newly launched tool designed to automate the process of improving agents by analyzing production traces, identifying patterns in failures, diagnosing root causes, and proposing fixes to prevent regressions. By clustering failures into named issues and suggesting targeted code changes, LangSmith Engine simplifies the traditionally manual agent development cycle, which involves tracing, identifying patterns, and creating ground truth datasets. The Engine proposes automated actions such as drafting pull requests, creating custom online evaluators, and adding examples to offline evaluation suites, thereby enhancing eval coverage and reducing the need for manual intervention. With its ability to monitor explicit errors, trace anomalies, and negative user feedback, LangSmith Engine aims to streamline issue resolution and improve agent robustness. It is currently available in public beta, with companies like Cogent, Harmonic, and Campfire already utilizing it to efficiently address issues and minimize triage time.

LangSmith Sandboxes have been announced as generally available, offering secure and scalable environments tailored for agent code execution, integrated with the Deep Agents SDK and LangSmith platform. Each sandbox operates as a hardware-virtualized microVM, ensuring isolation from other services and sandboxes, and supports any framework or custom code using the same SDK and API key. The need for these sandboxes arises from the growing use of agents requiring code execution as part of their workflow, necessitating a computer-like environment with strong isolation to mitigate risks such as supply-chain attacks. Traditional environments like containers may not provide adequate isolation for these dynamic, stateful agent workloads. LangSmith Sandboxes address these concerns by providing ephemeral microVMs that offer a filesystem, shell, package manager, and network boundary, while also supporting production controls, parallel workloads, and enterprise security features. These sandboxes enable agents to execute code, install dependencies, run tests, and maintain long-running sessions securely, enhancing capabilities for platforms like monday.com by allowing advanced workflows and richer functionalities for AI assistants like Sidekick. Future developments include local-to-cloud agent transitions, shared volumes for collaborative work, and enhanced execution tracing.

Managed Deep Agents, introduced in private beta, offers an API-first hosted runtime for creating and operating deep agents, allowing developers to focus on agent behavior rather than infrastructure. Built on the open-source Deep Agents harness, this solution provides a durable home for agents in LangSmith, facilitating programmatic creation, management, and execution of agents without the need for custom server setups. The system supports durable threads, streaming runs, checkpointing, human-in-the-loop workflows, and sandbox-backed execution, enabling agents to maintain context and improve over time. LangSmith's API enables developers to integrate tools and sandboxes, trace runs, debug behavior, and observe improvements, providing deep agents with the runtime necessary for long-term, tool-using, and context-preserving operations. This managed approach is suitable for various applications, including support, research, coding, data analysis, and internal operations agents, which require robust, durable execution capabilities to function effectively over extended periods.

SmithDB is a newly launched distributed database specifically designed to enhance agent observability for LangSmith, offering significant performance improvements by handling complex, large-scale trace data generated by modern AI agents. Built with Rust and utilizing the Apache DataFusion query engine, SmithDB supports advanced query patterns necessary for analyzing intricate agent behaviors, such as random access, interactive filtering, and full-text search, while maintaining low latency. Its architecture comprises object storage for trace data, a Postgres metastore for metadata management, and stateless services for ingestion and query processing, allowing it to efficiently manage workloads and scale across self-hosted and multi-cloud environments. SmithDB's implementation of a log-structured merge tree (LSM) and time-tiered compaction strategy optimizes for both write latency and query efficiency, while its innovative approach to handling large, unbounded payloads through late materialization and custom inverted indexing ensures fast query performance. The database has already been adopted by major clients like Clay, Vanta, Unify, and Cogent, who have noted significant improvements in speed and usability when analyzing large projects, making it a critical component in the ongoing development and evaluation of AI agents.

Deep Agents, built on the LangGraph runtime, leverages DeltaChannels to efficiently manage agent state checkpointing, significantly reducing storage requirements for long-running agents. Traditional checkpointing methods, which serialize a complete state snapshot at every step, lead to quadratic growth in storage, especially as agents accumulate extensive message histories and context. DeltaChannels, a new LangGraph primitive introduced in version 1.2, addresses this by storing only the incremental changes (deltas) at each step, while periodic full snapshots ensure manageable recovery costs. This innovative approach reduces storage from 5.3 GB to 129 MB for a coding agent running 200 turns, representing a 41× reduction. The introduction of DeltaChannels allows Deep Agents to maintain a comprehensive history of agent progress with minimal storage impact, making it an ideal solution for modern, long-running agents that require deep context management.

Organizations that excel at agent development have established a systematic lifecycle that allows them to ship agents early, learn from real-world usage, and iterate quickly. This lifecycle, comprising building, testing, deploying, and monitoring, ensures agents are not treated as one-off projects but are instead continuously improved through a repeatable process. The build phase involves selecting appropriate tools and environments, whether code-first or no-code, to create agents with varying complexity. Testing occurs before deployment, using evaluation datasets and simulations to ensure agents are ready for production. Deployment requires robust environments that support agents' long-running tasks and state management, including sandboxes and context hubs for versioning non-code elements. Monitoring focuses on understanding real-world agent behavior through traces and signals, linking feedback to specific actions for continuous improvement. Iteration is driven by rapid feedback loops, enabling teams to refine agents systematically and avoid relying on guesswork. Governance is crucial for managing costs, tool access, and discoverability of reusable assets, ensuring that fast iteration is possible without sacrificing control or consistency. By investing in shared infrastructure, organizations can make the agent development lifecycle an operational practice, transforming agents from demos into reliable production systems.

Automating company research through agentic orchestration and structured web intelligence can enhance due diligence processes in financial services by utilizing LangChain's Deep Agents for orchestrating tasks and Parallel's Task API for web research. This system streamlines the workflow of company investigations by assigning specific subagents to handle different research tracks such as corporate profiles, financial health, litigation and regulatory issues, news and reputation, and competitive landscapes. Each subagent is equipped to perform detailed investigations and produce structured intelligence reports with citations and confidence scores for each finding. The orchestrator coordinates these efforts, ensuring comprehensive coverage and cross-referencing findings to identify contradictions or low-confidence data, with the ability to adapt the research plan as new information emerges. This approach not only accelerates the due diligence process but also ensures compliance and auditability, as each claim is linked back to a primary source, meeting regulatory expectations for transparency and reliability. The entire process, validated through a case study on Rivian Automotive, leverages advanced AI and web research tools to produce reliable and verifiable company intelligence reports efficiently.

Agent observability is often initially perceived as a debugging tool, but its true potential lies in facilitating learning and improvement across the entire agent system. This process requires not only traces, which document what an agent did, but also feedback, which provides context and evaluation of the agent's actions. Feedback can come from direct user interactions, indirect user behaviors, or automated evaluations such as rules and LLM-as-judge systems. Effective agent observability platforms must store and integrate traces and feedback, allowing teams to analyze and enhance models, harnesses, and contexts. By doing so, observability transitions from simply recording actions to enabling systematic learning and development, transforming agent traces from mere logs into a comprehensive learning system.