A tool result does not always need to pass through the model

A conventional agent loads tool definitions into context, calls a tool, reads the result, and may copy that result into another call. Large catalogs and large intermediate documents make this expensive before the actual task becomes difficult.

Anthropic describes exposing tools through code interfaces instead. The agent discovers the relevant interface, then writes code that moves or filters data in the execution environment. A meeting transcript can travel from storage into a customer record without appearing twice in the model’s conversation.

The architectural distinction is between deciding what should happen and personally reading every byte involved. It can reduce context use and keep irrelevant data out of the model’s view. It also creates a real execution environment that needs its own permissions and controls; moving work into code does not remove responsibility for what that code can access.