July 2026 Summaries

Peter Steinberger's tweet sparked widespread discussion about the shift from loop to graph engineering in agent and workflow automation, highlighting the growing importance of graphs as a powerful data structure in computing. Graphs, used to represent information as interconnected nodes and edges, have become essential in various fields, as demonstrated by companies like Prefect and Dagster, which leverage graphs for orchestration and workflow automation. The conversation around directed agentic graphs emphasizes the balance between control and autonomy in agent behavior, where each node represents a unit of work and edges dictate the flow of tasks. This structure enables reproducibility, auditability, and security, particularly in complex workflows like refund processing, by allowing for specific control points and human interventions. Prefect is incorporating these concepts into its products, using directed agentic graphs to enhance the transparency and reliability of agent-driven operations, drawing on decades of orchestration experience while adapting to the unique challenges posed by AI. The discussion underscores the importance of adopting established best practices to avoid past mistakes and leverage the full potential of graph-based automation.

In episode two of the Fast MCP podcast, the concept of "context" in the realm of AI and its varied implications across different layers such as model, server, and application are dissected. Jeremiah Lowin, CEO, and Radhika Gulati, Product Marketing, delve into the complexities of context, explaining how it influences the behavior and decision-making of AI models by detailing the information available to them. The discussion highlights the challenges of managing context windows, which can overflow, leading to inefficiencies, and the evolution of context engineering as a practice to optimize the minimal yet effective delivery of context. The episode also addresses the confusion around the term "context" due to its multiple meanings in FastMCP, emphasizing the distinction between model context and server context. Additionally, the conversation touches on the development of MCP servers, the importance of progressive disclosure over the handshake problem, and how recent advancements such as tool search and code mode have made MCP servers more efficient. Finally, the podcast notes the upcoming changes with the MCP SDK version 2 and its implications for users, urging them to prepare for potential breaking changes and consider transitioning to FastMCP 4.

The author describes a transformative approach to managing sales tasks by utilizing an AI agent connected through the MCP protocol, which streamlines workflows by integrating various business systems into a single conversational interface. This setup allows for efficient management of tasks such as drafting follow-up emails, updating CRM data, and preparing for meetings, all without the need to manually switch between multiple applications. By implementing skills—customized procedures that automate repetitive tasks—the author has significantly reduced the time spent on non-selling activities, allowing more focus on customer interactions. The system is built on the Horizon platform, which centralizes permissions and access, ensuring security while enabling automation. Despite initial fears of complexity, the gradual development of this system has led to substantial time savings and improved productivity, with the author emphasizing the importance of starting small and building skills over time. This approach not only alleviates the administrative burden on salespeople but also enhances the quality and efficiency of their work.