SpatialClaw beats a prior spatial agent by 11.2 points on 20 tests

Spatial reasoning agents may need a better action interface more than a longer list of tools. NVIDIA AI wrote on X that “Code is the right action interface” for these agents, pointing to SpatialClaw, a training-free system that lets a VLM-backed agent write Python inside a persistent kernel. Instead of dispatching only fixed tool calls, the agent can compose perception modules, inspect intermediate outputs, and revise its strategy step by step.

The linked project page gives the strongest evidence. SpatialClaw reports an 11.2-point margin over a recent prior spatial agent across 20 benchmarks, with no benchmark-specific or model-specific tuning. It improves on 19 of 20 benchmarks on the same backbone and shows consistent gains across six VLM backbones. The page also reports an average +6.5 point gain over a no-tool baseline, with larger single-benchmark jumps such as DSI-Bench +17.6 points, MindCube +15.3 points, and MMSI +13.4 points.

NVIDIA AI’s account typically posts research, developer tooling, and infrastructure updates, and this item is more architectural than promotional. The claim is not that a new model alone solved spatial reasoning, but that executable code lets the agent turn perception outputs into reusable variables and computations. What to watch next is whether this pattern survives outside curated benchmarks: sandboxing, tool-state reproducibility, latency, and error recovery will decide whether code-as-action becomes a common interface for visual agents. The source tweet is available on X.