Stop Streaming Tools Through Your LLM

The 2026 shift from tool-calling to Code Mode: let the agent write one script instead of narrating fifty tool calls — and watch context tokens drop ~99%. TL;DR: The classic agent loop loads every tool...