S-Agent 2.0 teaches a coding agent to understand 3D and beats SpatialClaw by 7.8 points

The authors of S-Agent 2.0 took an off-the-shelf general-purpose coding agent, gave it a command line with a set of 3D perception tools, and got a notable jump on spatial benchmarks. The open Qwen3.8-27B reached 66.5 on MMSI-Bench and 82.3 on ReVSI. That is 13.9 and 20.5 points higher than the same agent without the tools. Gemini-3.8-Flash with the same toolset scored 74.1 and 90.5. The specialized SpatialClaw system, running on the same models, scored 62.6 and 74.5 with Qwen and 67.0 and 79.7 with Gemini.

Where accuracy was getting lost

Spatial reasoning requires a model to work out where objects sit in a three-dimensional scene, how they are positioned relative to each other, and how they move. Agentic VLMs have already improved noticeably on such tasks: instead of answering in a single pass, they gather data themselves through 3D perception tools. But the approach has a weak spot. On many questions the tools return correct data and the model still answers wrongly. Raw readings do not turn into a usable conclusion, and the authors are trying to close exactly this gap between the model and the tools.

Three steps instead of a pile of utilities

S-Agent 2.0 rests on one principle: help the VLM close the evidence loop. It has three parts.

• Getting scene data through a set of terminal commands.
• Storing findings in persistent scene memory, so the agent does not recompute the same things.
• Describing each reading in plain spatial language rather than bare numbers.

The system is not written from scratch. It is an ordinary coding agent that keeps all its flexibility: it can write scripts, combine calls, and check intermediate results. At the same time, every reading it relies on stays tied to the geometry the tools produced. The model does not invent distances. It reads them from the tool and receives an explanation of what they mean.

The numbers and the comparison with SpatialClaw

• Qwen3.8-27B: 66.5 on MMSI-Bench and 82.3 on ReVSI. The gain from the tools was +13.9 and +20.5. Judging by these figures, the baseline without tools is about 52.6 and 61.8.
• Gemini-3.8-Flash: 74.1 and 90.5, a gain of +11.2 and +9.6.
• SpatialClaw on Qwen3.8-27B: 62.6 and 74.5. S-Agent 2.0 is ahead by 3.9 and 7.8 points.
• SpatialClaw on Gemini-3.8-Flash: 67.0 and 79.7. S-Agent 2.0 leads by 7.1 and 10.8 points.

The authors themselves point out an important caveat. The stronger the base model, the smaller the gain from the tools. But the gain shrinks differently in the two systems. SpatialClaw's gain on ReVSI drops from +13.2 to +2.7 when moving from Qwen to Gemini, so it nearly vanishes. For S-Agent 2.0 on the same move, it falls from +20.5 to +9.6 and stays substantial. In practice this matters more than absolute scores: if a tool stops helping as models get stronger, you will have to throw it out within a couple of releases.

The authors also note separately that S-Agent 2.0 works in different coding-agent harnesses, including Codex and Claude Code. So the result is not tied to a single product.

What you can take into your own work

The main takeaway for people building agent systems concerns the interface. Quality depends not only on the set of tools but also on how their outputs are returned to the model. Scene memory and verbal descriptions of readings can carry over to other tasks: robotics, video analysis, navigation. If you are building an agent on a ready-made coding harness, it makes sense to test the "commands, memory, interpretation" scheme on your own data without rewriting the harness itself.

The limitations also deserve attention. All the figures come from two benchmarks, and the gain was not measured on real cameras or noisy sensors. For now this is a strong result for a research setting, not a ready-made system for a robot.