Post

1/6 People often say LLM agents are “just another abstraction layer.” I think that confuses abstraction with automation. A useful tool is not automatically an abstraction.

2/6 Abstraction is not just hiding complexity. It is building a model that lets us safely ignore details because the behaviour we depend on is stable enough to reason about.

3/6 A good abstraction: - hides irrelevant details - preserves important semantics - exposes a clear contract - reduces knowledge consumers must carry - makes change safer

4/6 Abstraction is not: - automation - delegation - orchestration - compression - a black box - “it usually works” Those things can be useful. But usefulness is not abstraction

5/6 If a tool forces you to understand its hidden quirks, model-specific behaviours, prompt sensitivities, or implementation details to use it safely, then it has not fully abstracted them away. That is closer to a leaky abstraction.