AWS open-sources Strands Decider 2B and Strands Box for agent development

The release gives teams a small local decision model and a sandbox for controlling agent actions more tightly.

AWS has open-sourced two additions to its Strands agent stack: Strands Decider 2B, a small model for making structured choices inside agent workflows, and Strands Box, a developer-preview sandbox for enforcing execution policy. For teams building agentic systems, the point is to separate decision-making from runtime control so routing, tool selection, and guardrails do not depend only on prompts. AWS says Strands Decider 2B is built from Qwen3.5-2B, uses a pointer head and rank-16 LoRA tuning, and ships with weights, training data, and scripts so it can run locally. The company reports a median decision latency of about 115 milliseconds on an NVIDIA RTX 3090 and 153 milliseconds on an Apple M3 MacBook for small tasks, plus 72.3% accuracy on 231 public tasks in its JevBench v19 evaluation. Strands Box combines operating-system isolation with Dogwood policies to govern shell, Python, MCP tools, filesystem actions, and network access, including rules such as requiring tests before a push or limiting API calls.