Building internal tools across multiple code repositories creates a specific problem for AI coding agents: they make locally reasonable decisions that are wrong at the system level. An agent editing a Go service has no idea that a TypeScript frontend depends on the exact API shape it just changed.
Superblocks, the internal tools platform, published a detailed breakdown of how they solved this with what they call a "workspace repository" - a single orchestration repo that contains zero application code, just infrastructure glue. It uses a repos.yaml manifest as the source of truth for all repository URLs and branches, Git worktrees for isolated branch work without copying files, and structured AGENTS.md files that give AI agents visibility into the full system architecture.
The practical result: when an AI agent works inside this workspace, it can see how services connect to each other before making changes. Cross-repo pull requests become visible to one another instead of existing in isolation.
This is an engineering blog post, not a product launch. Superblocks is sharing internal practices, not selling a new tool. But the pattern they describe - giving AI agents a structured map of your entire system before letting them write code - is one that any team running AI coding assistants across multiple repos should study. The specific tools they use (Tilt for service orchestration, Nix for environment management, just as a command runner) are less important than the principle: AI agents need cross-service context to avoid creating bugs that only show up at integration time.