Spatial reasoning asks a model to determine where objects lie in 3D space, how they relate, and how they move. Recent agentic vision‑language models (VLMs) improve on these tasks by letting the model actively collect evidence with 3D perception tools instead of answering in a single pass. Yet many questions still receive correct raw readings from the tools while the model fails to draw the right conclusion, exposing a gap between tool output and usable inference.
We therefore introduce S‑Agent 2.0. It equips an off‑the‑shelf general coding agent with a command‑line perception toolkit and designs the interaction around a single principle: close the evidence loop. The agent acquires scene evidence, stores it in a persistent scene memory, and interprets each reading in plain spatial language. This preserves the full flexibility of a coding agent while keeping every operation grounded in the geometry supplied by the tools.
On the MMSI‑Bench and ReVSI benchmarks, the toolkit lifts the open‑source model Qwen3.8‑27B by +13.9 and +20.5 points respectively, and the closed‑source model Gemini‑3.8‑Flash by +11.2 and +9.6 points. With S‑Agent 2.0, Qwen3.8‑27B reaches 66.5 on MMSI‑Bench and 82.3 on ReVSI, surpassing the specialized system SpatialClaw (62.6 and 74.5) on the same backbone. Gemini‑3.8‑Flash attains 74.1 and 90.5, also beating its baseline scores of 67.0 and 79.7.
As backbone models become more capable, the gains from tools diminish; however, SpatialClaw’s gain on ReVSI drops from +13.2 to +2.7, while ours remains at +9.6. S‑Agent 2.0 also transfers smoothly to other general coding agent frameworks such as Codex and Claude Code, demonstrating strong versatility.
Review