December 2025 Summaries

Fastweb + Vodafone, part of the Swisscom Group, is transforming its customer service operations in Italy by implementing AI-driven solutions to handle the complexities of telecommunications support. The company has introduced two key projects: Super TOBi and Super Agent, leveraging LangChain and LangGraph technologies. Super TOBi enhances customer interactions by using a structured graph-based decision-making flow to resolve inquiries efficiently, achieving a 90% correctness rate and an 82% resolution rate. Super Agent, designed for internal use, empowers call center consultants with precise diagnostics and guidance, contributing to One-Call Resolution rates above 86%. Both initiatives rely on LangGraph's graph-based orchestration and Neo4j's knowledge graph to streamline processes and ensure consistent, context-aware responses. Continuous improvement is facilitated through LangSmith monitoring, offering insights into system performance and enabling rapid adjustments. As Fastweb + Vodafone expands these capabilities, they aim to further innovate in the telecommunications AI space while delivering exceptional customer experiences.

Polly is an AI-powered assistant integrated into LangSmith that aids in debugging, analyzing, and improving agent performance, addressing challenges encountered by developers when building agents. Unlike simple language model calls, agents involve complex interactions with longer prompts, extended traces, and multiple turns, making it difficult to identify issues manually. Polly excels in this environment by understanding agent architectures, recognizing failure patterns, and offering insights through natural language queries, effectively serving as an expert agent engineer. It assists in debugging individual traces, analyzing entire conversations, and engineering better prompts, streamlining the process of pinpointing inefficiencies and refining agent behavior. Polly's capabilities are powered by LangSmith's comprehensive tracing infrastructure, which captures detailed agent activity, enabling Polly to provide actionable insights and improvements. Now available in beta, Polly represents a significant advancement in agent engineering, promising to enhance the efficiency and reliability of agents developed on the LangSmith platform.

LangSmith Fetch is a newly launched CLI tool that integrates LangSmith tracing capabilities directly into developers' terminals and IDEs, enhancing workflow efficiency by eliminating the need to switch between interfaces. Designed for developers who prefer command-line environments, it allows seamless access to trace data without navigating the LangSmith UI, making it especially useful for those using coding tools like Claude Code or Cursor. LangSmith Fetch supports two primary workflows: immediate access to recent traces for debugging and bulk export of data for evaluation or analysis, all through simple command-line operations. By providing programmatic access to agent execution data, it enables coding agents to perform advanced debugging tasks, such as analyzing decision points and suggesting improvements. The tool is compatible with existing LangSmith setups, requiring no additional configuration, and offers flexibility beyond real-time request/response patterns, making it a versatile addition to a developer's toolkit. Available on PyPI, LangSmith Fetch is positioned as a fundamental building block that enhances agent observability and debugging capabilities directly from the terminal.

Debugging deep agents presents unique challenges compared to simpler LLM applications due to their complexity, which involves longer prompts, extensive traces, and multiple user interactions. LangSmith addresses these challenges by offering tools specifically designed for debugging deep agents, such as the AI assistant Polly and the LangSmith Fetch CLI. Polly helps analyze trace and thread data, suggesting improvements and identifying inefficiencies, while the Fetch CLI provides seamless integration with coding environments, allowing the retrieval and analysis of execution data. This approach helps developers gain insights into agent behavior, improve system prompts, and enhance the overall performance of deep agents. The LangSmith platform emphasizes the importance of tracing to capture comprehensive execution data, enabling the use of AI for effective debugging and optimization of deep agent workflows.

Agent engineering is an emerging discipline that focuses on refining non-deterministic language model systems into reliable production experiences, responding to the unpredictability inherent in agents that can handle complex tasks across multiple steps. Unlike traditional software development, where inputs and outputs are largely known, agent engineering must contend with the vast and unpredictable range of user inputs in natural language, necessitating a continuous cycle of building, testing, observing, refining, and repeating. Successful companies like Clay, Vanta, LinkedIn, and Cloudflare are pioneering this approach, integrating product thinking, engineering, and data science to create systems that adapt and reason in real-time. This new engineering standard involves a shift from exhaustive pre-launch testing to using production as a primary learning environment, where each iteration provides insights into user interactions and system reliability. The rise of agent engineering is driven by the increasing capability of language models to perform sophisticated workflows, demanding a new approach to ensure these systems are trustworthy and effective in real-world applications.

DeepAgents CLI is an open-source, Python-based terminal coding agent built on the Deep Agents SDK, designed for tasks across diverse domains such as software engineering, biology, security, and gaming. It includes features like shell command execution, file operations, web search, task planning, and persistent memory storage. To evaluate its performance, it was tested on Terminal Bench 2.0, a benchmark encompassing 89 tasks, where it achieved a mean score of 42.65%, comparable to other implementations using the same model. The testing process is facilitated by Harbor, a framework that executes agents in containerized environments, ensuring isolated and clean evaluations through sandboxing. DeepAgents Harbor allows for scalable evaluation using multiple sandbox providers like Docker and Daytona. The benchmark tests involve a range of tasks, from simple to complex, and results are verified automatically with reward scoring. The evaluation underscores DeepAgents CLI as a competitive solution, with future plans to enhance performance through systematic analyses and optimizations.

LangChain recently developed and deployed four applications using their Deep Agents harness, including a coding agent, LangSmith Assist, a personal email assistant, and a no-code agent building platform. These applications necessitated the creation of evaluation patterns specific to Deep Agents, which involve bespoke test logic for each data point due to the unique success criteria for each instance. Evaluations can be conducted through single-step tests to validate immediate decision-making, full agent turns to assess the complete execution, and multi-turn tests to simulate extensive user interactions, each requiring a clean environment for reproducibility. LangSmith's integrations facilitate these evaluations by allowing for detailed assertions on agent behavior, including trajectory, final responses, and other generated state, while also offering tools to handle complex evaluation environments efficiently.

LangSmith Agent Builder is an innovative tool that allows users to create production-ready agents without coding, offering a dynamic alternative to traditional AI workflow builders. Unlike workflows, which require a predetermined step-by-step approach, agents can autonomously determine and execute the necessary steps, adapt to new information, and improve through user feedback. This tool simplifies the process of building agents by using chat-based guidance to help users create detailed prompts and select the appropriate tools, thereby enabling anyone to develop agents in minutes. The public beta release introduces new features like multi-model support, programmatic invocation, and simplified UI, along with capabilities to connect external APIs and internal systems, allowing teams to collaborate and customize agents efficiently. It addresses challenges such as the steep learning curve of AI tools and the need for secure, flexible customization, making it easier for teams to automate tasks like research, communication, and project management, ultimately saving time and increasing productivity.