Strands Decider 2B is a small model that chooses between options a developer supplies rather than writing long text. It covers model routing, tool selection, classification, guardrails and context management, and returns a decision with confidence scores. It is built on Qwen3.5-2B with a pointer head in place of the language-model head, and AWS has released its weights, training data and scripts for local use.
AWS reports a median decision latency of about 115 milliseconds on an NVIDIA RTX 3090 and 153 milliseconds for small tasks on an Apple M3 MacBook. On JevBench, it reports 72.3% accuracy across 231 public tasks for its v19 checkpoint. The package is named strands-decider and can be used from its command-line interface or inside Strands agents.
Strands Box is in developer preview and pairs operating-system-level isolation with Dogwood, an open-source policy language that defines which actions an agent may take. It covers shell commands, Python, MCP tools, filesystem operations and network access, and policies can allow or deny actions based on their content and earlier steps. AWS says it works with different models and frameworks, locally or in the cloud, and its policies run outside the model rather than relying on prompts.
Why it matters if you build agents
Teams that route tools or classify inputs can test a small local model as an alternative to a general-purpose model call, though early GitHub discussions raise issues with concurrent requests, Apple Silicon and cold-start latency. Strands Box lets teams enforce rules such as requiring passing tests before a code push, without replacing their existing agent architecture. Because enforcement sits outside the prompt, it applies even when a model ignores its instructions.
Get stories like this every weekday
One short email with the day's agent tooling news.