July 2025 Summaries

AI coding tools are revolutionizing software development by enabling developers to generate code rapidly using large language models (LLMs), a process sometimes referred to as "vibe coding." While this method can enhance productivity by allowing developers to create code intuitively without writing every line, it also poses challenges such as logic errors, security vulnerabilities, and technical debt due to insufficient understanding of the generated code. To address these issues, tools like Graphite are transforming vibe coding into responsible AI-assisted development by integrating high-quality, human-in-the-loop feedback cycles. AI-powered code reviewers, such as Graphite Agent, are becoming necessary to manage the increasing volume of AI-generated code by quickly identifying logic errors, security gaps, and performance issues, thus allowing human reviewers to focus on broader architectural decisions. By adopting AI in the code review process, teams can achieve faster iterations and higher-quality software, ensuring both efficiency and innovation without compromising on code quality.

Anthropic's Claude Code has seen a 300% increase in usage over the past three months due to its potential to automate tasks typically sidelined by engineering teams. While initial tests showed promise in identifying bugs and staging fixes, challenges arose when it generated large, unreadable blocks of code, creating difficulties in review and implementation. To address these issues, the team experimented with stacking pull requests (PRs) to break down large code changes into smaller, manageable tasks. This approach was implemented using the GT MCP tool, which organizes tasks into a sequence of focused PRs, mimicking the workflow of senior engineers. This method has enhanced the review process, making AI-generated code more manageable and useful in real engineering work, with the team successfully building features, resolving bugs, and adding tests across their codebase. The GT MCP is now available for broader community use, promising efficiency gains by helping AI agents to stack PRs effectively.