I Built an AI System for 80+ Microservices. Six Months Later, My Whole Team Uses It.

In April, the author shared a system they built on Claude Code that could turn an epic into PR-ready code across 80+ microservices. However, the author realized that their system had flaws and decided to make improvements. One of the biggest issues was duplicated rules across 18 different files, which caused drift and inconsistent behavior.

To address this, the author consolidated all the rules into a single file with short IDs, making it easier to manage and reference the rules. The author also separated rules into two categories: judgment calls and invariants (mandatory rules). They implemented a permission system that blocked commands that violated the invariants, providing a stronger wall against rule violations.

Additionally, the author introduced read-only agents for functionality such as log searching, MongoDB querying, SQL execution, and metric reading. These agents did not have write access, reducing the chances of accidental changes. The author also started creating runbooks for common incidents, which helped new joiners quickly find solutions.

After implementing these changes, the author noticed that a few teammates started contributing to the system by adding features like local service clusters, database agents, and log agents. This change made the system more collaborative and valuable to the entire tech team, as new joiners could now rely on the shared knowledge base instead of constantly reaching out to the author for help.