July 2025 Summaries

The text discusses the concept of "deep agents," which are advanced AI models capable of handling complex tasks over extended periods, as opposed to "shallow" agents that lack depth in planning and execution. Deep agents, exemplified by applications like Claude Code, Manus, and Deep Research, utilize a detailed system prompt, a planning tool, sub-agents, and a file system to achieve this depth. These components enable the agents to manage context effectively, divide tasks among sub-agents, and utilize a file system for memory and collaboration. The text also introduces an open-source package called "deepagents" designed to facilitate the creation of custom deep agents for various verticals, incorporating key characteristics inspired by successful models like Claude Code. This package allows users to easily build tailored deep agents by providing customizable prompts, tools, and sub-agents, showcasing its application in a simple "deep research" agent example.

LangSmith has introduced Align Evals, a new feature designed to improve the alignment of evaluator scores with human preferences in application development, particularly when using language models as judges. Inspired by Eugene Yan's work, this feature allows LangSmith Cloud users, and soon LangSmith Self-Hosted users, to calibrate evaluators to better reflect human judgment through an interactive interface that facilitates prompt iteration and side-by-side comparisons of human and AI-generated scores. Align Evals addresses the challenge of evaluator score discrepancies by providing tools to identify unaligned cases, establish a baseline alignment score, and iteratively refine evaluator prompts to achieve better alignment. The feature enables developers to select evaluation criteria, create representative data sets for human review, assign expected scores, and test LLM evaluator prompts against these benchmarks. Future enhancements include performance analytics and automatic prompt optimization, aiding developers in building more effective evaluators.

Bertelsmann, one of the largest media companies globally, has developed the Bertelsmann Content Search system to address the challenges of navigating its vast and decentralized content ecosystem. Utilizing LangGraph, a multi-agent system, the company has transformed content discovery by orchestrating searches across various platforms without centralizing all data into a single system. This innovative approach employs specialized agents for different content domains, routing queries intelligently to provide coherent insights and enhance creative productivity. The system enables faster content discovery, cross-platform insights, and democratized access, allowing creative teams to focus more on creation rather than searching, thus fostering collaboration and agility across the organization. The successful implementation of this technology demonstrates the potential of AI in media, offering a scalable and flexible solution for managing and utilizing extensive content portfolios.

Over the past year, agentic applications, such as workflow copilots and codegen assistants, have emerged, transforming how we conceptualize applications by integrating tool use, memory, and reasoning to complete complex tasks. These agents, unlike traditional apps, act independently, remember past interactions, and collaborate, necessitating a new infrastructure beyond existing serverless or microservice architectures. The LangGraph Platform is introduced as a solution, providing the necessary agent infrastructure to support durable execution, complex state management, human-in-the-loop coordination, and handling of bursty concurrency and streaming. This platform enables agents to run for extended periods, manage state efficiently, and handle traffic surges, while also allowing for intermediate output streaming to enhance user and developer experiences. By offering these capabilities, LangGraph Platform aims to simplify the deployment and management of agentic workloads, allowing developers to focus on agent logic without the burden of building the supporting infrastructure.

Open deep research has emerged as a significant application in the realm of agent-based systems, gaining popularity with products from companies like OpenAI, Anthropic, Perplexity, and Google, as well as numerous open-source implementations. It involves a flexible, multi-agent approach to research that adapts its strategies based on the user's request, enabling detailed and comprehensive reports by leveraging various models, search tools, and MCP servers. The process features three main phases: scoping to gather necessary context, research using a supervisor and sub-agents to delve into sub-topics, and report writing to compile the findings into a coherent deliverable. The system's architecture allows parallelized research efforts, ensuring depth and efficiency, and emphasizes context engineering to manage token usage and enhance performance. The open deep research platform, built on LangGraph, encourages experimentation and customization, offering users the opportunity to test and deploy their agents via the Open Agent Platform.

LangSmith Deployment, previously known as LangGraph Platform, is now available in the AI Agents and Tools category of AWS Marketplace, allowing customers to leverage AWS accounts for streamlined discovery, purchasing, and deployment, thus enhancing the development of agentic workflows. LangSmith provides enterprise teams with capabilities for prompt engineering, evaluation, and observability, enabling structured prompt development, performance evaluation, and deep visibility into AI applications. Meanwhile, LangGraph Platform offers infrastructure for deploying and managing stateful, long-running agents with features like visual workflow development and centralized agent management. Through AWS Marketplace, teams can expedite procurement, maintaining control over licensing and compliance while ensuring data privacy by running the platforms within their AWS VPCs. LangChain, the open-source framework behind these offerings, supports building critical AI applications and is widely adopted by leading engineering teams globally.

As businesses increasingly explore building AI agents, this guide provides a structured approach for developing such agents from concept to impact, using the example of building an email agent. It outlines a six-step framework starting with defining the agent's job through realistic examples, followed by designing a detailed operating procedure to ensure the problem scope is manageable. The process continues with building a minimum viable product (MVP) focusing on the most critical reasoning tasks, and then connecting the MVP to real data using orchestration logic. The guide emphasizes the importance of testing and iterating to refine the agent's performance before deployment, and it highlights the necessity of continuous monitoring and adaptation based on user feedback and real-world usage patterns. Ultimately, the guide stresses that launching an agent is just the start of an iterative process to ensure the agent remains useful, reliable, and aligned with user needs.

Context engineering is an essential discipline for enhancing the performance of AI agents, focusing on effectively managing the limited context window of language models (LLMs) by employing strategies such as writing, selecting, compressing, and isolating context. This approach is crucial for agents that interleave LLM calls with tool usage, often in tasks that require long-running interactions and significant memory management. Effective context engineering can address issues like context poisoning, distraction, confusion, and clash by ensuring that only relevant information is retained or recalled. Tools like LangGraph and LangSmith offer frameworks to support these strategies, providing features like state management, sandboxing, and multi-agent architectures to optimize context usage and improve agent performance. By mastering context engineering, developers can create AI systems that are more efficient, scalable, and capable of handling complex tasks through judicious use of memory, tool integration, and feedback loops.