Drew Breunig
AI practice & strategy
Writer on AI, data and strategy. Author of widely-cited pieces on agentic coding practice.
Recent activity
-
The Pace Layers framework, introduced by Stewart Brand, organizes fields and categories by their rate of change, with faster layers influencing slower ones, and can be applied to the AI ecosystem to understand its dynamics. The framework proposes six macro layers, with the fastest layers (days) at the top and slowest layers (decades) at the bottom, and highlights the need for balance between the layers to maintain resilience. By applying Pace Layers to the AI ecosystem, developers can better understand the relationships between the different layers and the potential consequences of rapid change on the system as a whole. AI summary
Read more → -
The plain-English prompt that makes prototypes effortless turns out to be a poor way to specify how a system should behave, and the bill arrives slowly, disguised as ordinary progress, until the application can barely move.
Read more → -
Developers can leverage AI skills that double as documentation, as they often provide better explanations than traditional documentation, and can be iterated upon more easily. To create effective human-centric documentation, developers should focus on building mental models that teach the art of using tools, rather than providing exhaustive details, and prioritize clarity over completeness. AI tools like scaffold-docs can assist in this process by providing a structured framework and style guidelines. AI summary
Read more → -
Overfitting models to their specific harness designs can lead to reduced generalizability, making 3rd party harnesses less valuable when used with frontier lab models, which are then locked into their pre-trained behavior. This trend may result in frontier models resembling appliances rather than general platforms. As a result, application building may become easier for enterprises, but at the cost of increased lock-in. AI summary
Read more → -
The following is a write up of a talk I delivered at MLOps Community’s “Coding Agents” conference, on March 3rd. There’s a video version of the talk available on YouTube.
Read more → -
The development of software has evolved into three models: the Cathedral, the Bazaar, and the Winchester Mystery House, characterized by careful planning and management (Cathedral), open and community-driven development (Bazaar), and idiosyncratic, sprawling, and cobbled-together software (Mystery House), enabled by the abundance of cheap code. AI summary
Read more → -
The 2nd phase of agentic development in open source coding projects is characterized by reimaginings of existing software, rather than clones or ports, leveraging the low cost of code and the capabilities of coding agents and LLMs to rapidly test, patch, and improve new tools. This phase is marked by projects such as Cheng Lou's TypeScript library for text layout and Cloudflare's EmDash, a modern CMS, which reimagined the job to be done without relying on existing standards. AI summary
Read more → -
The article visualizes the components used to assemble a system prompt in Claude Code, a large company app, including always-included components (e.g., intro, output style), conditional components (e.g., system rules, task execution guidelines), and variations within those components. The visualization highlights the complexity of context engineering and the importance of harnesses in assembling dynamic system prompts. AI summary
Read more → -
AI systems like Anthropic's Mythos have become so capable at detecting and exploiting vulnerabilities that security spending is now comparable to the number of tokens used to find exploits, effectively rendering traditional notions of security and cost less relevant. This trend echoes the proof-of-work system in cryptocurrency, where success is tied to raw computational work. As a result, open-source software and code review processes are becoming increasingly important for securing systems, with the cost of hardening a system now being limited by the market value of an exploit rather than the budget. AI summary
Read more → -
When implementing agentic coding, it's essential to implement to learn, rebuild often, and invest in end-to-end tests to uncover decisions and improve specifications. Agentic coding benefits from frequent rebuilds, where code is cheap, allowing developers to reinvent and compound their efforts, and tests should measure the product's functions rather than its performance. Developers should also focus on the hard, difficult work of intuitive design, performance, security, resilience, and systemic architecture, and minimize easy tasks to optimize time for these critical areas. AI summary
Read more →