AWS launches Strands Decider 2B and Strands Box for AI agents

AWS released two open-source projects: Strands Decider 2B, a 2B-parameter model that picks between predefined options in agent workflows, and Strands Box, a sandbox with Dogwood policies for controlling agent execution. The model runs at ~115 ms median latency on an RTX 3090 and scores 72.3% on JevBench.

• Strands Decider 2B: 2B parameters, built on Qwen3.5-2B with a rank-16 LoRA adapter
• Median decision latency ~115 ms on RTX 3090, 153 ms on Apple M3
• 72.3% accuracy across 231 public JevBench tasks (v19 checkpoint)
• Strands Box in developer preview: kernel isolation plus Dogwood policies