Coding agents have become very good at writing code to solve a specific, well-defined problem. Give an agent a clear task and a test to pass, and it will usually produce working code quickly.
Where engineering still struggles is at the system level. Human judgement is the real job now: deciding how components should interact, where boundaries belong, and what the system should look like months from now. Most of that work happens outside the context window.
An agent trying to understand a whole codebase by reading raw source files hits a wall fast. Context windows fill up with implementation details, and the agent loses sight of the bigger picture. We need a simpler abstraction that lets agents grasp the broader architecture of a system without reading every line of code.
Google’s Goldfish-Elephant model suggests the same idea. It separates long-horizon design work from fresh, zero-context review and execution instead of keeping everything in one long chat session. That only works when the system structure is captured cleanly outside the conversation.
If you follow a microservice architecture, FINOS CALM (Common Architecture Language Model) is one project to look at. It lets you define services, interfaces, and controls as machine-readable JSON alongside your code, giving both humans and agents a simple map of how the system fits together.