August 2025 Summaries

Erik Munson, the founding engineer of Day AI, has a rich history of managing orchestration challenges across various companies, including HubSpot and Netflix, and now applies his expertise in building an AI-native CRM. His journey began at HubSpot, where he helped develop a CRM product that significantly contributed to the company's growth. During his tenure, Munson tackled challenges related to real-time data processing and orchestration, utilizing tools like Kafka and Redis to handle high volumes of email tracking data. At Netflix, he further honed his skills while working on Spinnaker, an open-source CI/CD tool, which reinforced his understanding of the complexities of workflow orchestration. At Day AI, Munson faced unprecedented data stream processing challenges, opting for Inngest as the technical stack to manage real-time data from various communication platforms efficiently. Inngest's event-driven architecture allowed for simplified workflow management without extensive infrastructure, enabling Day AI to maintain high reliability and operational efficiency with a small team. This choice marked a pivotal moment in Munson's career, as it provided a generalized solution to the orchestration challenges he encountered, allowing him to focus on product development rather than operational complexities.

Over the past year, Context Engineering has emerged as a crucial concept in the development of AI systems, surpassing the initial prominence of Prompt Engineering. While Prompt Engineering focused on crafting the right prompts for language models, Context Engineering emphasizes providing AI agents with an optimal set of tools, memory, and data to build and manage their context effectively. This shift underscores the importance of context quality, which is vital for the success or failure of AI agents. Context Engineering involves structuring workflows for data gathering and classification, as well as short- and long-term memory retrieval, aiming for a robust orchestration layer that connects LLM applications to diverse data sources. It parallels Software Engineering principles in its architectural approach, offering flexibility in integrating tools and systems to enhance agent capabilities. This new approach allows AI agents to integrate deeply with users' environments, exemplified by AI applications like Day AI and Windmill, which utilize advanced Context Engineering architectures to offer tailored, autonomous assistance.

A small team building a resume parser for a job-matching platform initially chooses Kafka for its speed and low latency but soon becomes overwhelmed by the complexity of maintaining the infrastructure required to handle retries, deduplication, and rate limiting. As the team struggles with managing race conditions, custom coordination logic, and Kafka infrastructure, they discover Inngest's Flow Control, which offers built-in functionalities such as debouncing, throttling, prioritization, and retries without the need for additional glue code or infrastructure management. This shift allows the team to focus more on product development, reducing bugs and API usage, while enhancing their ability to implement new features rapidly. Inngest's Flow Control simplifies the process by providing declarative primitives, reducing the overhead associated with Kafka and Redis, and allowing the team to deliver features more efficiently.

Inngest developed a robust multi-agent system using their developer-first framework, AgentKit, to efficiently query and analyze 3.5 million powerlifting records in real-time, addressing challenges like state management and context maintenance. The system allows users to ask natural language questions about powerlifting meets and receive structured results by orchestrating three specialized agents: a Routing Agent to determine database access needs, a Query Agent to convert language into structured queries, and a Summary Agent to interpret results. The agents operate within a carefully structured architecture that leverages a language-to-SQL pattern, using ClickHouse for data storage and precise query handling, with the agents defined and tested in isolation before integration. This approach provides clear separation between intent parsing and data summarization, facilitating accurate and deterministic data access while allowing for fully inspectable logic. AgentKit's capabilities enable the orchestration of these agents as functions within a single network, demonstrating significant improvements in control, accuracy, and efficiency in handling large structured datasets.