One tool to rule them all: AI agents ditch their toolboxes for a single code slot
Philipp Schmid writes that AI agents are collapsing their whole tool catalog into one exec(code) tool, known as Code Mode. The model writes a short TypeScript or Python script that chains and filters actions inside a sandbox, rather than making one tool call per turn. Only a brief result goes back into the model's context. Schmid says Factory's Droid 1.0 gained 3.5 points on Terminal-Bench 4 and cut model spend by 15% after switching. He cautions that not every agent will end up with just one tool.
Why it matters: It suggests AI agents may work more efficiently by writing short scripts instead of calling many separate tools, which could improve performance and lower costs for developers and users, though Schmid says one tool won't suit every agent.
Comments
No comments yet.