# Code as the Action Interface for Multimodal Agents

Vision-language and multimodal agents increasingly use executable code as their action space — writing and running programs to compose perception steps, call tools, and inspect intermediate results — as a training-free path to more reliable spatial and visual reasoning; expect more frameworks built on this pattern.

- Conviction: 30 / 100 (weakening)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-07-02T00:00:00.000Z
- Last updated: 2026-07-07T14:00:02.329Z
- Canonical: https://polylog.news/ai/trends/code-as-action-multimodal-agents
- Publisher: Polylog
- Affected regions: Global

## Recent score history

- 2026-07-06: 31
- 2026-07-07: 30

## Recent evidence

- [confirms] NVIDIA's SpatialClaw Uses Code as the Action Interface for Spatial Reasoning (2026-07-02): NVIDIA's SpatialClaw is a training-free framework in which a vision-language agent writes and runs Python to compose perception steps and inspect intermediate results for spatial reasoning.