While loops with tool calls

Daniel Leitnack and Chris Benson engage with Jared Zonaryk on the evolution of prompting in AI, particularly focusing on advancements in reasoning models from OpenAI. They discuss "context engineering," which emphasizes the significance of the information included in prompts and how users can now mix and match inputs more effectively. The conversation highlights the transition from complex workflows to more autonomous interactions, resembling while loops with tool calls, reflecting the rapid evolution in prompting techniques.

Jared shares insights on the need for continuous updates in training materials due to AI advancements and the challenges of keeping pace with technological changes. He introduces two philosophical approaches to large language models (LLMs): an academic understanding of their context and a practical "tinker and builder" philosophy that focuses on matching inputs to outputs. The importance of testing effective prompts is emphasized, as many users can create automations without needing to understand complex concepts like self-attention.

The discussion also covers prompt sensitivity and the impact of formatting changes on outputs. Jared warns that excessive academic knowledge can hinder practical application, introducing the concept of "LLM idiom" to describe the nuances of interacting with language models. He notes that structured prompts work well for precise tasks but may not suit creative endeavors, highlighting the need for context in prompting.

The unpredictability of AI models, especially in customer support applications, is addressed, with structured paths developed to guide user interactions. Recent advancements have improved models' instruction-following capabilities and their ability to produce structured outputs. Tool calling has become essential, enabling prompts to access various functions, such as issuing refunds or checking user status. The emergence of autonomous agents, like Claude and Codex, has been supported by simplifying processes, allowing models to operate more freely.

User responsibilities in this evolving landscape are discussed, emphasizing that both developers and end-users can expect improved performance as models explore solutions more effectively. The aim is to create AI agents that behave more like humans, transitioning from rigid structures to flexible interactions. Jared shares an example of an agent created using cloud code to update documentation daily, illustrating the rapid development of effective solutions.

Testing strategies for AI workflows are explored, including unit testing input-output functions and the challenges posed by recursive functions and loops. The concept of "agent smell" is introduced to identify potential performance issues, with suggested metrics for evaluation. The conversation highlights the high failure rate of AI pilots and the importance of effective tools and collaboration between engineering teams and domain experts to ensure success in AI applications.

Jared advocates for a gradual approach to AI development, emphasizing the importance of starting with small, achievable goals and iterative testing. He illustrates the need for collaboration between engineers and domain experts, where engineers focus on product shipping and domain experts provide insights on quality and relevance. Looking ahead, Jared expresses excitement about the evolution of coding agents as headless tools that can be integrated into workflows, democratizing coding for those without a computer science background.