June 2026 Summaries

In the second installment of a series on creating a fully automated cloud software factory, the focus is on incorporating spec-driven development for addressing complex or ambiguous issues that cannot be resolved with straightforward implementation. The process begins with a Triage agent assessing issues and applying labels based on their complexity and alignment with the project’s roadmap and vision, triggering a Spec agent if needed. This agent generates detailed product and technical specifications stored under a "specs" directory linked to the issue number, which guide the subsequent implementation phase. The framework leverages existing infrastructure, such as Docker images and GitHub actions, to facilitate this workflow, and encourages interactive refinement of specifications to ensure alignment with project goals. The ultimate aim is to establish a system that can efficiently handle, spec, and implement issues, with future posts set to explore quality assurance through review and verification agents. Readers are invited to try the setup using Warp’s cloud platform and provide feedback on their experience.

The post outlines the creation of a "cloud software factory" designed to automate parts of the software development lifecycle (SDLC) using agents that perform tasks such as issue triage, specification writing, implementation, and verification. The initial setup, demonstrated using the Warp platform, focuses on automating the triage and implementation stages by using coding agents like Warp and cloud runners such as Oz. The process involves setting up a repository on GitHub, configuring Docker images, and using GitHub actions to invoke skills that manage issue labeling and resolution. The approach aims to decouple from specific platforms, promoting flexibility and scalability, with future posts promising to expand functionality to include spec-driven development and more human-in-the-loop processes. The post provides resources and instructions for setting up this automation in a repository, advocating for a gradual build-up of a comprehensive automation system.

/pr-walkthrough is a tool designed to enhance understanding of pull requests (PRs) by generating an interactive website that visualizes key components, such as system overview, data flow, code dependency graph, and user experience, using D3 for visualization. This tool can be easily integrated into continuous integration (CI) processes and hosted using GitHub Actions and GitHub Pages, allowing users to explore each PR with a custom graphical user interface. The creator encourages feedback and contributions to improve the tool, highlighting its potential to streamline PR reviews by providing a comprehensive and interactive examination of changes.

The post explores the concept of creating automated feedback loops to enhance the performance of Skills, using a specific example of replatforming websites from WYSIWYG platforms to self-hosted code on Vercel. It introduces an "observer" Skill that evaluates the effectiveness of the replatforming process by running tests, identifying behavioral and visual differences, and suggesting improvements based on structured data and state-of-the-art models. The observer Skill is designed to optimize costs while maintaining quality by analyzing patterns and proposing updates to the underlying "inner" Skill, thus continuously improving it through iterative loops. The process is facilitated by tools like Warp and Oz, which support orchestration of multiple agents, and the entire approach is open-sourced for broader application. Although the method is not flawless and has limitations like the risk of finding local maxima, it offers a practical way to ensure Skills operate efficiently by continually refining them. The post also mentions a new podcast called Talking Slop, along with related articles discussing broader applications of Skills and loops in software development.

Warp is transitioning towards an era of Automated Development, where engineers focus on creating internal systems, or "cloud software factories," that autonomously build products. Instead of directly coding, the goal is to develop a system that automates as many tasks as possible, with success measured by the percentage of changes shipped automatically and the associated costs. This shift requires a new mindset where interactive agent involvement is seen as a failure to learn from, pushing towards full automation. Warp aims to refine its processes by continuously evaluating and improving its automation systems, with the ultimate aim of achieving self-improving factories. This approach aligns with their mission to empower developers to deliver superior software more swiftly, as demonstrated by their open-source Warp terminal and Oz platform initiatives.

In recent discussions, the concept of "loops" has been highlighted as a method to enhance the capabilities of agents through self-improvement cycles. A practical example involves using cloud agents and Skills to create self-improvement loops, where agents refine their Skills based on external feedback. Specifically, the process is demonstrated with an issue triage Skill that categorizes issues into buckets like "ready-to-implement" or "needs-info." This system involves an inner agent loop that applies the Skill and records interactions, and an outer agent loop that evaluates these runs and modifies the Skill based on feedback, such as changes made by human reviewers. This feedback mechanism is facilitated through Warp and Oz's cloud agent platform, integrating with GitHub's issue tracking, allowing for continuous improvement with each iteration. The framework behind this approach is used to manage the Warp open-source repository and is available for others to adopt.

Rectangle Health, a healthcare payment technology company, has significantly enhanced its software development process by integrating a custom AI Teammate named Rex, powered by Oz, a cloud orchestration platform. Initially a personal project by Senior Software Engineer Matthew Chastain, Rex evolved into a pivotal tool for the company, capable of writing 54% of its own code and managing tasks across multiple repositories. This innovation addresses the challenge of scaling productivity and reducing tech debt by using specialized agents that automate workflows from ticket creation to pull requests, all initiated through Slack. In just 60 days, Rex contributed 139 commits and 97,000 lines of code, becoming essential to Rectangle Health's engineering operations. The platform has cultivated a culture of autonomous, efficient development, with plans for further integration with AWS Bedrock and the establishment of self-hosted infrastructure. The success of Rex has inspired other teams within Rectangle Health to explore building similar agents on the Oz platform, highlighting a shift towards an agentic development approach.

To ensure the successful completion of a project by an agent, it's essential to write detailed specifications, including product specs that describe user behavior and tech specs that outline the implementation strategy. These specifications should be documented as markdown files and included in the implementation pull request for team review. The process involves several steps: creating a product spec to define a feature from the user's perspective, followed by a tech spec to detail the implementation plan. After these specs are created, the agent is tasked with implementation, which must be validated to ensure adherence to the specs. An additional validation step involves using specific skills in Warp to check for consistency between the specs and the implementation, and using Oz for UX validation. These skills are available on GitHub for adaptation and integration into projects, facilitating both agent and human collaboration in the development process.

In a recent overhaul of their marketing site, the team moved away from visual editors, which, despite their speed, had led to technical and design issues, and instead embraced a code-driven approach to enhance workflow efficiency and agent collaboration. Recognizing that the problem was rooted in workflow rather than just website design, they focused on creating a system that integrates seamlessly with agents, utilizing well-documented frameworks such as Next.js and Tailwind, and separating content management into a platform like Sanity to allow agents to operate more autonomously. This shift allows agents to handle operational tasks while humans maintain oversight and strategic direction, thus fostering a more dynamic and scalable system where changes are transparent, reviewable, and reversible. By building a modular component architecture, the team ensured that agents could work effectively within a structured environment, optimizing for flexibility and coherence at scale while maintaining the human touch in defining intent and making critical decisions. The transformation signifies a broader change in design philosophy, highlighting the importance of building adaptable systems that keep pace with rapid change, enabling both agents and humans to contribute to the creation and management of digital assets efficiently.