February 9, 2026

A few weeks ago, I was staring at a system prompt that was 127,000 tokens long. One hundred and twenty-seven thousand tokens - just the system prompt. Most of it was, you guessed it, tool schemas: JSON definitions for every possible thing the agent could do, most of which it would never touch in any given run. My solution? I propose AgentScript: a code-first execution language where the LLM can only interact with the outside world by writing programs.