Abstractions

Abstractions let AI agents work at the right level without losing the forest for the trees.

What an Abstraction Is

An abstraction is an interface that hides implementation details. A file system interface is an abstraction—you can swap in a mock, network storage, or in-memory implementation without changing the code that uses it.

This isn’t just good software design. It’s a way to scope what the AI can touch.

Why This Helps AI

When an agent tries to do everything at once, it gets bogged down in low-level details. The context fills up with implementation noise, and the model loses track of the high-level goal.

Abstractions let you split the work:

1. One agent creates the abstraction — builds the library, defines the interface, handles the low-level complexity
2. Another agent uses it — works at the high level, focused on the application logic

Together they achieve more than one agent trying to do both.

Examples

Extracted libraries: You build a reusable component library in one session. Future sessions import those components without needing to understand their internals. The agent thinks about what to render, not how rendering works.

Config-driven modules: You write a dynamic core once, then let agents work only with JSON config files. They can configure behavior without touching the engine that interprets the config.

Interfaces and mocks: Define a data access interface. The agent writes business logic against that interface. The actual database implementation is abstracted away—the agent doesn’t need to know about connection pooling or query optimization.

The Pattern

Extract complexity into stable abstractions. Let agents work at higher levels of abstraction where they can stay focused on the immediate goal.