PubNub has transitioned from using ad-hoc prompting with Claude Code to implementing a structured, multi-agent subagent workflow to enhance the reliability and efficiency of software development processes. This approach utilizes specialized subagents that handle distinct tasks such as product specification, architecture review, implementation, and testing, each with its own system prompt and tool permissions. The workflow is designed to improve reproducibility, separation of concerns, and governance by codifying steps and using hooks to manage transitions between tasks while keeping humans in the loop for oversight and approval. Subagents are defined with Markdown and YAML, and can be discovered and invoked by Claude Code, allowing for the integration of other language models like GPT-5 for additional review stages. By refining and customizing subagents over time, PubNub aims to maintain an efficient, secure, and scalable engineering system, facilitating faster onboarding and safer automation.