July 2026 Summaries

The exploration of harness tuning for open models, specifically using Nemotron 3 Ultra within Deep Agents, highlights the importance of aligning model capabilities with the environment they operate in to maximize performance. While open models like Nemotron 3 Ultra offer cost-effective and adaptable alternatives to frontier models, their effectiveness heavily depends on the compatibility of the harness they are paired with. The study demonstrates that a well-tuned harness, which includes elements such as system prompts, tool descriptions, and middleware, can significantly enhance model performance without altering the model itself. This iterative tuning process involves evaluating model actions, diagnosing behavioral patterns, and making targeted adjustments to improve outcomes. The research underscores that while harness tuning can significantly enhance model performance in specific tasks, it has limitations and cannot compensate for inherent model deficiencies that require post-training solutions. The case study of Nemotron 3 Ultra showed improvements in task performance such as summarization and tool use, indicating that a properly tuned harness can lead to substantial cost savings while maintaining high-quality outputs, though it also highlighted that some long-term behavioral improvements would require changes at the model level rather than through harness adjustments.

The newly announced NemoClaw for LangChain Deep Agents blueprint, developed in collaboration with NVIDIA, aims to enhance the performance of production agents by providing an open and governed framework for enterprises. This blueprint integrates LangChain Deep Agents Code, NVIDIA Nemotron 3 Ultra, and NVIDIA OpenShell runtime to allow teams to customize, secure, and optimize agents for specific workloads, achieving advanced performance at a significantly reduced cost. By tuning the model, harness, evaluations, and runtime together, enterprises can build proprietary systems that reflect their unique domain expertise while maintaining control over the agent stack. The open model layer of Nemotron 3 Ultra is complemented by a tuned agent harness from LangChain and a secure runtime from NVIDIA OpenShell, enabling efficient and transparent deployment of agents. This approach allows companies to fine-tune model weights, customize harnesses, and control runtime based on various requirements, ultimately leading to more efficient and specialized agents. Supported by an ecosystem of partners like EY, Baseten, Fireworks, and Nebius, the blueprint facilitates the deployment of open agentic models in production, ensuring transparency, auditability, and cost-efficiency while meeting enterprise standards.

Deep Agents Code, when run as a NemoClaw blueprint, provides enterprises the ability to utilize coding agents for modernizing legacy systems while maintaining control over sensitive codebases. This approach leverages NVIDIA's open-source technologies, including Nemotron 3 Ultra and OpenShell, to create a secure sandbox environment where coding agents can operate with autonomy but without risk of unauthorized actions. The system is designed to allow enterprises to govern their infrastructure, ensuring code and data residency remain within chosen boundaries. It offers a structured process for modernizing legacy applications by mapping dependencies, refactoring code incrementally, and maintaining an audit trail through all stages of the transformation. This method aims to address the challenges of updating critical systems, such as COBOL-based business logic, by providing a controlled, auditable, and vendor-independent framework for code modernization.

Continual Learning, Harness Engineering, and Post-Training focus on curating data at scale to enhance and improve AI agents by running experiments. This approach was discussed at the AI Engineer World Fair, where the importance of data mining from Traces was highlighted as a crucial tool for companies to understand and improve their agents. Continual Learning involves agents acting in their environment and reintegrating the information gained back into the system, akin to human learning. Traces, which are projections of agent experiences, serve as valuable data to mine for understanding agent behavior. As agents become more complex and produce larger volumes of data, specialized systems like LangSmith Engine have been developed to efficiently process and analyze these traces, finding signals and issues, generating code fixes, and storing crucial information. The integration of open models, which are cost-effective and intelligent, allows for better processing of this data. A practical recipe for agent improvement involves using a combination of Harness Engineering and Fine-Tuning, allowing teams to iteratively enhance agent performance through data collection, evaluation, and continuous experimentation. This iterative process is essential for adapting to increasing data production and enhancing agent capabilities over time.

Schneider Electric is leveraging artificial intelligence to enhance energy efficiency and sustainability across industries, with a focus on electrification, automation, and digitalization. The company operates an extensive AI program through its AI Hub, which involves 350 experts deploying over 60 AI agents to optimize energy consumption, prolong asset lifespans, and boost developer productivity. Central to their strategy is the use of AI to forecast energy demand and production, enabling users to shift electricity usage to cost-effective, eco-friendly times. Schneider's AI operations are underpinned by a robust LLMOps framework built around the LangSmith and LangChain ecosystems, which supports observability, evaluation, and deployment of AI products. This framework ensures data privacy, compliance, and high-quality agent performance, fostering a collaborative environment where subject matter experts can contribute to the development and refinement of AI solutions. Schneider's AI initiatives, such as the internal AI Assistant "One Jo" and the Customer Success Manager Copilot, demonstrate the company's commitment to integrating AI in critical infrastructure while maintaining rigorous cybersecurity standards. Through these efforts, Schneider is advancing its mission to drive sustainable energy management and industrial automation, with a vision of significantly reducing global energy consumption and carbon emissions.

LangSmith addresses the challenges faced by teams using multiple coding agents, such as the increased costs and fragmented data visibility that result from "tokenmaxxing," where excessive spending is mistakenly equated with productivity. The solution involves consolidating data from various agents like Claude Code, Codex, Cursor, and GitHub Copilot Chat into a unified trace model, enabling teams to see and compare expenses across tools in a consistent format. This visibility allows for optimization, where inefficiencies are identified and actionable recommendations are provided by the Engine feature to refine workflows. Additionally, the LLM Gateway offers governance by capping costs at user, team, and organizational levels, and it can integrate open-source models for cost-effective alternatives in appropriate scenarios. By providing a single platform for monitoring, debugging, and measuring coding sessions, LangSmith empowers teams to manage their AI tool usage effectively, ensuring that spending aligns with actual value delivered.

Recursive Language Models (RLMs), introduced by Alex Zhang and MIT CSAIL researchers, aim to counteract the issue of context rot in language models by utilizing programmatic orchestration and dynamic subagents. Unlike traditional models that may struggle with context accumulation over extended sequences, RLMs operate by running code in a REPL (Read-Eval-Print Loop) environment, allowing them to dispatch subagents and recursively process input context. This approach enhances performance by splitting tasks into manageable units and orchestrating them through code rather than relying solely on the model's judgment. Deep Agents, a platform that incorporates RLM support, leverages dynamic subagents and a lightweight code interpreter to facilitate complex workflows across a mix of models, optimizing for tasks like classification and data aggregation. Benchmarking against the OOLONG dataset demonstrates that RLM-enabled agents outperform standard models, particularly in scenarios requiring long-context reasoning, despite higher latency and token costs. By enabling models to write recursive loops for task-specific contexts, RLMs and dynamic subagents provide a structured method to enhance the reliability and scalability of automated workflows.

Pendo's Novus is a product agent designed to automatically detect usability issues in live applications, promptly fix the underlying code, and enhance the overall user experience, achieving a 90%+ success rate in evaluations. The integration of LangSmith tracing has been pivotal, allowing for detailed monitoring and debugging of Novus in production by providing comprehensive trace views that reveal user interactions and system behaviors. Novus utilizes product analytics and session replays to identify and address actionable issues, correlating user behavior with specific code files and suggesting fixes, thereby closing the feedback loop between developers and product managers. This approach allows product teams to maintain high shipping velocity without sacrificing quality, as Novus automatically corrects issues before they become significant problems. The use of LangSmith for trace tagging and cost monitoring ensures efficient resource allocation and insights into customer usage, enabling tailored product improvements and outreach strategies.

OpenWiki is an open-source tool designed to automate the generation and maintenance of documentation for codebases, enhancing coding agents' understanding of repositories. By creating a dynamic wiki linked to a coding agent, OpenWiki ensures documentation remains current as code changes, addressing the challenge of outdated documentation in large repositories with frequent pull requests. Inspired by previous codebase wiki initiatives, OpenWiki integrates with agent instruction files rather than storing extensive documentation within them, allowing agents to access the necessary context efficiently. It supports various model providers and utilizes GitHub Actions to schedule regular updates, ensuring the wiki reflects the latest code changes. OpenWiki aims to streamline the documentation process, reducing the manual effort required from developers and potentially expanding its application to other workflows beyond coding.