Best Practices with Claude Code Subagents Part II: Moving from Prompts to Pipelines

Between August 2025 and the present, significant advancements were made in Claude Code subagents, transitioning them from "structured prompts" to a more functional runtime environment. These changes include the introduction of forked contexts, which allow subagents to operate in isolated sessions, preventing "context rot" and maintaining clarity. The concept of "Skills" has been refined to unify user and model-invoked capabilities, enhancing composability and efficiency. Backgrounding long tasks has improved workflow fluidity, enabling agents to perform time-consuming operations without disrupting the main thread. The document structure has evolved, emphasizing a "Plan → Execute → Verify" pipeline, with clear directories and versioned contracts for planning, while the CLAUDE.md file now serves as a "Project Constitution" with concise directives. Hooks have been expanded to improve governance, offering pre-tool use validation and post-tool logging, and MCP server integration has been streamlined for tool discovery and management. These enhancements aim to create a more resilient, organized, and efficient pipeline for teams, reducing chaos and improving productivity within the Claude Code framework.