AWS has introduced two open-source components for developers building agents: Strands Decider 2B, a small model that chooses among predefined options, and Strands Box, an isolated execution environment with policy controls. They separate the decisions an agent makes from the boundaries imposed on its actions.
Decider 2B is based on Qwen3.5-2B and replaces the usual language-model output layer with a pointer mechanism that scores supplied choices. It targets model routing, tool selection, classification, guardrails and context management. AWS released the weights, training data and scripts. The company reports median decision latency near 115 milliseconds on an Nvidia RTX 3090 and 153 milliseconds for small tasks on an Apple M3 MacBook, but those figures are vendor tests.
Strands Box, currently in developer preview, routes shell commands, Python, tool calls, filesystem operations and network access through kernel-enforced isolation and the Dogwood policy language. A rule can require tests before a code push or limit external API calls based on earlier actions.
The controls operate outside the language model, reducing dependence on prompts for security. Developers still need to test production fit: early discussions have identified concurrency, Apple Silicon and cold-start issues, and AWS’s published Decider benchmark does not establish an advantage over competing decision models.