How We Taught AI Agents to See the Bigger Picture

The problem: Good code is not always accepted code AI agents can write working code. They can make changes that compile, pass tests, and finish tasks. And yet the pull request might still be rejected. That happens because every codebase has its own internal logic. Teams have naming conventions, preferred APIs, and patterns they trust. Some of those rules are documented, but many are not. Pavel Sher (Technical Lead for TeamCity) has reviewed thousands of commits over the years. That experience helps him spot when a change technically works but still does not fit the project. He can suggest better naming, a more appropriate approach, or a solution that stays consistent with the rest of the codebase. This led us to a practical question: how can we help AI agents see the bigger picture, too? The trap: Agents learn from what they see When you ask an AI agent to write code, it usually looks through the codebase for examples. It tries to infer conventions from existing code and follow the pat