Code-as-action vs structured tool-calls — SpatialClaw explained
What is a code-as-action interface? An AI agent acts by writing and running executable code — one Python cell per step against a stateful kernel — instead of calling a fixed menu of tools.
SpatialClaw gives a vision-language agent this interface for 3D/4D spatial reasoning: write one cell, run it, read the result, then write the next. Training-free, it averages 59.9% across 20 benchmarks — +11.2 points over the prior spatial agent — across six VLM backbones.
Full explainer (interactive): https://learnaivisually.com/g/spatial...
Source: https://arxiv.org/abs/2606.13673
Learn AI & GPUs visually — free interactive courses at learnaivisually.com
#AIagents #CodeAct #LLM #AI