S-Agent 2.0: Coding agent outperforms specialists in spatial reasoning

S-Agent 2.0 is a spatial reasoning agent built on the ready-made OpenHands coding agent, featuring CLI perception tools and persistent scene memory. With a frozen model stack and Gemini-3.8-Flash backend, it achieves 74.1% on MMSI-Bench and 90.5% on ReVSI, outperforming SpatialClaw and GPT-6.

Compiled by Sergey KostenchukPublished 2026-10-22Updated 2026-10-22

2026-10-22 Coding Google

🛠 S-Agent 2.0: Coding agent outperforms specialists in spatial reasoning

The agent is built on OpenHands with CLI perception tools: it independently takes 3D measurements, stores them in scene memory, and is required to describe measurements in words rather than relying on intuition. The stack—Depth Anything 3, SAM 3, GeoCalib, Pi3-MOS—is frozen; with a Gemini-3.8-Flash backend, it achieves 74.1% on MMSI-Bench and 90.5% on ReVSI, compared to SpatialClaw (67.0%, 79.7%) and GPT-6 (71.0%, 84.2%).

🌍 The improvement came not from a new controller, but from a measurement protocol: spatial intelligence is embedded into standard agent pipelines without a specialized system.

👤 Code is open under Apache 2.0: recon, segment, distance, compass, bev commands, and benchmarks are available on GitHub. For VLM agents and robotics, this is a takeaway: the model references calculated numbers rather than guessing. Caveat: two benchmarks are insufficient to dismiss the need for specialization.

Source 1: https://arxiv.org/abs/2610.28354 Source 2: https://github.com/Ropedia/S-Agent-2.0

Sources