Teams building AI agents against large API surfaces can adopt a code-generation interface (e.g., two `search`/`execute` tool calls) to slash context token usage by orders of magnitude and unlock native programming constructs like loops and parallelization that JSON tool calling cannot efficiently express.