A practical article on AI software development reveals how one team registered more than 20 code repositories, pulled only those involved in each delivery, and isolated changes with Git worktrees. We see AI programming competition shifting from “can it write code?” to “can it control cross-repository engineering?”
What this is
The core of this approach is “one main repo plus multiple business repos.” The main repo holds workflows, standards, scripts, and a repository manifest; the Agent (an AI assistant capable of understanding and executing development tasks) starts from the main repo. For repositories requiring changes, it creates worktrees (independent working directories provided by Git) from a fixed baseline, preventing code changes from directly contaminating the original project directories.
It also requires the Agent to know before every operation which repositories are readable and writable, as well as the commands for compilation, testing, and code checks. Any action outside that scope is rejected outright by command-line tools. The article also openly identifies two pitfalls: worktrees can fill up disk space, and unmerged changes can be lost.
Industry view
The value of this approach lies not in putting all code back into a monolithic repository, but in preserving enterprises’ existing multi-repo structure while adding permissions, baselines, and integration rules for AI. For teams with complex legacy systems, this is more realistic than migrating repositories and easier to adopt incrementally.
But we do not think it is a universal answer. The objections are direct: a main repo may encourage AI to access code beyond its boundaries, while its manifest may fall out of sync with actual permissions. Worktrees also introduce additional disk, cleanup, and merge costs. More importantly, Git operations and build checks can constrain the process but cannot ensure that cross-repo changes are functionally correct; test coverage and human review remain indispensable.
Impact on regular people
For enterprise IT: AI development tools are entering the permission-governance phase. Companies need to map repositories, test commands, and merge processes first; otherwise, expanding automation scope could make management more costly than manual development.
For individual careers: Programmers’ core work will increasingly shift toward defining requirements, setting constraints, and specifying acceptance outcomes. Those familiar with Git, testing, and architectural boundaries will be better positioned to turn AI output into reliable delivery.
For the consumer market: Multi-platform products may update faster, but consumers will not immediately notice changes in AI programming tools. The first beneficiaries will be enterprise software and development services with complex internal systems and frequent iteration cycles.