Compound Engineering Plugin formalizes agent loop

This is less a “plugin” than a process layer for AI coding teams. The bet is that better software comes from better loops, not just better prompts.

• –It turns the common agent workflow into explicit stages with specialized commands and agents, which should reduce improvisation and make results more repeatable
• –The knowledge-capture piece is the real moat: storing patterns in repo docs and `CLAUDE.md` means future runs can compound instead of relearning the same mistakes
• –Cross-tool support for Claude Code, Codex, Cursor, and others makes it more of an ecosystem play than a single-client extension
• –The tradeoff is overhead: teams need to buy into planning, review, and documentation discipline for the gains to show up
• –For agent-heavy codebases, this looks like a practical attempt to make “autonomous” coding less chaotic and more auditable