November 2025 Summaries

Anthropic has introduced the concept of agent skills, which are essentially folders containing a SKILL.md file and associated resources that allow agents to perform tasks more effectively by dynamically loading necessary information. This approach has been integrated into the deepagents-CLI, an open-source coding assistant, which allows agents to utilize a collection of public skills by accessing the local filesystem. Generalist agents like Claude Code and Manus demonstrate that with access to basic computing tools like bash and filesystem operations, they can accomplish a wide range of tasks with minimal tools. The skills approach offers two main advantages over traditional methods: it improves token efficiency by loading only necessary information when required, and it reduces cognitive load by simplifying the toolset needed for task execution. Additionally, skills facilitate continuous learning as agents can develop and share new skills, enhancing their ability to adapt to novel tasks.

Deep agents, equipped with filesystem tools, can effectively manage and utilize large sets of context, addressing challenges related to context retrieval and storage in AI models. By using filesystems, agents can optimize context engineering by efficiently storing, retrieving, and updating necessary information, thus preventing issues like retrieving too many tokens or lacking critical context. This capability allows agents to intelligently search for niche information and learn over time by incorporating user feedback into their filesystem. Filesystems provide a flexible interface for agents to dynamically handle vast amounts of data, enhancing their ability to answer questions accurately and maintain relevant context across interactions. This approach also opens up possibilities for agents to grow their skillsets and instructions, making them more adept in future interactions. The Deep Agents open-source repository exemplifies these practices, allowing developers to leverage filesystem capabilities in building more reliable and context-aware agents.

Jimdo, a leading European website builder, has expanded its services over 18 years to support self-employed entrepreneurs with tools beyond website hosting, including AI-driven features for bookings, analytics, transactions, and marketing. Recognizing its customers' challenges in managing various business functions without dedicated teams, Jimdo developed Jimdo Companion, an AI-powered business advisor using the LangChain ecosystem. This platform provides personalized guidance and integrates real-time data analysis to enhance business performance, offering features like Companion Assistant and Companion Dashboard to help users optimize their operations. Jimdo's use of LangChain.js and LangGraph.js facilitates seamless integration of data sources and AI capabilities, while LangSmith ensures monitoring and quality assurance, leading to significant improvements in customer engagement and business outcomes. With a focus on transforming customer interactions into business value, Jimdo aims to evolve into an AI-powered business platform, enhancing efficiency and effectiveness for solopreneurs.

ServiceNow is enhancing its sales and customer success operations by developing an intelligent multi-agent system using LangChain's LangSmith and LangGraph tools, which address the challenges of agent orchestration and observability. This system orchestrates the entire customer journey, from lead identification to post-sales adoption and expansion, by coordinating complex workflows that were previously fragmented. ServiceNow's approach involves using a supervisor agent for orchestration and specialized subagents for specific tasks, activated by customer signals and lifecycle stages. LangGraph provides the necessary tools for sophisticated multi-agent coordination, enabling a modular and efficient development process, while LangSmith offers detailed tracing and evaluation capabilities to improve agent performance. The evaluation framework includes custom metrics tailored to each agent's tasks, leveraging LLM-as-a-judge evaluators and ensuring regression prevention. Currently in testing, ServiceNow plans to integrate real user data and utilize new multi-turn evaluation features for more comprehensive assessments of agent performance across the entire customer interaction.

Vivek Trivedy introduces Sandboxes for DeepAgents, a new feature enabling the secure execution of arbitrary DeepAgent code in remote environments provided by partners like Runloop, Daytona, and Modal. Sandboxes offer a safe, configurable space to run code, protect local machines from potentially harmful operations, ensure clean environments with specific dependencies, allow parallel execution, support long-running tasks, and guarantee reproducibility across teams. The integration process involves setting up the sandbox, executing commands remotely, and maintaining visibility into sandbox operations. Users can start by creating accounts with partner platforms, storing necessary API keys, and using the DeepAgents CLI for easy sandbox setup. Despite the isolated environment, precautions against prompt injection and secret management are advised. The initiative aims to enhance development flexibility, with plans to expand configuration options and examples for practical integration.

AWS re:Invent attendees interested in agent development can visit Booth #524 at the Venetian Expo Center from December 1-4 to explore the latest features like Insights Agent and Multi-turn Evaluations, with the opportunity to engage in technical discussions and demos with the engineering team. Notable events include a LangChain x OpenSearch event on December 1, a Redis Partner Lightning Session on December 3, and AI LIVE at re:Invent with LangChain on December 2, which cover topics ranging from context engineering to the development of full-fledged agentic systems. Attendees can also learn about self-hosting LangSmith on AWS Marketplace for those exploring this option. The team is available for meetings to discuss challenges such as evaluation strategies and scaling multi-agent systems, and those unable to attend can connect afterward for further discussions.

Liam Bush outlines the development and optimization of an internal agent using LangChain, LangGraph, and LangSmith to improve technical support efficiency. Initially built as a prototype for product Q&A and customer demonstration, the LangChain Chatbot faced underutilization due to a lack of comprehensive answers for complex queries. The team observed a successful three-step manual workflow involving documentation, knowledge base, and codebase searches, which they then automated with a Deep Agent incorporating specialized subagents for each step. This agent synthesizes information into accurate, detailed responses, saving engineers significant time on debugging. Recognizing its effectiveness, the team applied similar strategies to the public Chat LangChain, enhancing it by integrating direct API access for documentation, knowledge bases, and codebases, rather than relying on vector embeddings. The system uses specialized tools to mirror human search behaviors, prompting agents to refine queries and focus on reasoning over retrieval. The implementation of production middleware ensures reliability and cost-efficiency, while future updates aim to broaden public access to codebase searches.