February 2026 Summaries

Developing a marketing strategy for exe.dev has been approached with a "show, don't tell" philosophy, leveraging the technical capabilities of the product to naturally attract interest and engagement. Despite the overwhelming marketing noise surrounding AI, the unique functionalities of exe.dev, such as the ability to write private apps from a phone using their agent Shelley, have proven effective in drawing users. The platform allows users to create and share apps effortlessly, exemplified by the author's experience of developing a network testing tool for exe.dev's global anycast network. This process highlighted the potential of using exe.dev for rapid app development without extensive code interaction, thereby creating a new paradigm for shareable, small-scale app creation that can be utilized for both professional and personal projects.

In reflecting on his experiences contributing to the Go project, the author highlights the value of code reviews as a powerful tool for operational learning, surpassing even the detailed commit history in providing insights into the decision-making process, mid-stream changes, and the underlying values driving decisions. The narrative explores the cognitive challenges of implementing new features without reviewing their implementations, leading to a loss of understanding and confidence in project direction. To address this, the author has developed a workflow where an agent reviews the code, and he subsequently reviews the review, allowing him to maintain an abstract understanding of the code and provide efficient direction. This process, which includes a detailed numbering system for comments, questions, and suggestions, ensures clarity and minimizes surprises, ultimately converging on essential commentary while discarding less relevant feedback. The approach underscores the significance of human involvement in code review processes and the continuous adaptation of workflows for optimal results.

At exe.dev, the team developed a link shortener named Slinky, which stands out due to its "Edit with Shelley" feature, allowing users to modify the link shortener's functionality directly. This feature enables users to incorporate template parameters into Slinky URLs, facilitating the creation of customizable links for specific use cases, such as Honeycomb queries. By clicking the "Edit with Shelley" button, users can access the Shelley agent on the same virtual machine to implement changes, treating the software like a wiki for seamless customization and enhancements.

Exe.dev has been active since its launch in late December, introducing several new features and enhancements. Among these are the fast "cp" command for creating virtual machine (VM) copies, a restart command in the SSH lobby to manage VMs, and the ability to rename VMs post-creation. The web UI now includes a billing portal for account management, and early work on larger disk support is underway. Users can now send and receive emails directly from their VMs, and DNS records are reliably created before command completion. The interface now supports dark mode, and the Shelley agent, which powers exe.dev's default image, has become open-source, with new features like custom model support and visualization capabilities. Shelley also supports skills, iframe displays for HTML output, and can self-upgrade from the UI. The team continues to encourage engagement through their blog and Discord community.

The text discusses the cost dynamics associated with using language model (LLM) tools, specifically focusing on how cache reads can become a significant expenditure during long conversations with coding agents. By analyzing various conversations, it reveals that by around 50,000 tokens, cache reads can dominate the costs, ultimately reaching up to 87% of the total cost in some cases. The discussion explains that while fewer LLM calls are cheaper, they may lead to inefficient navigation of tasks without feedback, posing a dilemma between cost-efficiency and task accuracy. Strategies like using subagents, starting new conversations for context management, and employing tools like "keyword search" are suggested to balance costs and maintain effective task management. The piece raises questions about whether cost, context, and agent orchestration issues are interconnected and whether approaches like Recursive Language Models could be beneficial, emphasizing ongoing considerations for developers working on platforms like exe.dev and Shelley.