What's Happening?
Robert Stark, CTO at Twnstr, has articulated a new paradigm for software development, emphasizing that 'code is grown, not created' in the age of AI. His essay, which gained traction on Hacker News, details how a year of AI-heavy development has reshaped
his approach to shipping software. Stark argues that human developers can no longer hold an entire system in memory for weeks, as AI agents accelerate codebase growth beyond individual comprehension. He advocates for a 'horticulture' approach to coding, focusing on planting constraints, inspecting growth, and pruning. Key practices include using 'constitution documents' for project principles, 'metaplans' for hierarchical goal setting, and enforcing 'domains and layers' for agent search efficiency. Stark also stresses the importance of a 'rewrite cadence' for cheap pivots and scheduled 'slop cleanup' to prevent technical debt.
Why It's Important?
Stark's insights are crucial for organizations grappling with the integration of AI into their software development workflows. His framework addresses the challenges of increased output from AI agents and the resulting strain on review bandwidth. By shifting from an 'authorship' to a 'horticulture' mindset, teams can better manage the rapid evolution of AI-generated code. The emphasis on 'constitution documents' and 'metaplans' provides a structured approach to guide AI agents, ensuring alignment with project goals and architectural principles. This can lead to faster prototyping, more efficient exploration of alternative designs, and a reduction in the 'verification tax' associated with re-deriving decisions. Ultimately, adopting these practices can help companies leverage AI to accelerate development cycles while maintaining code quality and architectural integrity, impacting the efficiency and cost-effectiveness of software engineering across industries.
What's Next?
Organizations are encouraged to implement Stark's recommendations immediately. This includes writing one-page 'constitutions' before AI agent sprints, creating 'metaplan' directories, and linking relevant parent files to every session. Weekly 30-minute 'slop passes' with metrics for lines removed and duplicate helpers merged are also advised. Additionally, teams should add architectural linters that AI agents cannot bypass and log rewrite experiments, even if not shipped, to capture learning. For structured training, resources like Claude for work workshops are suggested. Stark's playbook is most effective when a CTO has repo authority, but its principles can be adapted even when constitutions are missing or compliance requires human sign-off on every line of code. The ongoing discussion on Hacker News will likely provide further insights and practical applications from other practitioners.
Beyond the Headlines
Stark's vision of 'code is grown, not created' signifies a fundamental shift in the role of human developers. As AI agents become more sophisticated, the human role may evolve from direct code authorship to that of an architect, gardener, and quality controller. This raises ethical questions about accountability for AI-generated errors and the potential for 'false confidence' in AI summaries. The shift also implies a change in hiring signals, with 'fluency with agent interrogation' becoming as important as raw syntax speed. The concept of 'slop compounding' highlights the critical need for proactive cleanup and maintenance in AI-driven development, preventing the accumulation of technical debt that could cripple future projects. This new paradigm challenges traditional notions of software ownership and craftsmanship, pushing the industry towards a more collaborative, AI-augmented future.













