AWS has introduced two open-source projects for AI agent development: Strands Decider 2B, a small model designed to make decisions within agent workflows, and Strands Box, a sandbox for controlling agent execution.
AWS has introduced two open-source projects for AI agent development: Strands Decider 2B, a small model designed to make decisions within agent workflows, and Strands Box, a sandbox for controlling agent execution. Both address different aspects of building agents that use tools, execute code and interact with external systems, according to the company.
Strands Decider 2B is a two-billion-parameter model built to select between predefined options rather than generate lengthy text responses. It supports tasks including model routing, tool selection, classification, guardrails and context management. Instead of asking a general-purpose language model to generate an answer, developers can provide a set of choices and receive a decision with confidence scores.
The model is based on Qwen3.5-2B, with its language-model head replaced by a pointer head that scores candidate answers, and uses a rank-16 LoRA adapter for fine-tuning. AWS has released the model weights, training data and scripts, allowing developers to run and modify the model locally.
AWS reports a median decision latency of approximately 115 milliseconds on an NVIDIA RTX 3090 and 153 milliseconds for small tasks on an Apple M3 MacBook. The company evaluated the model using JevBench, reporting 72.3% accuracy across 231 public tasks for its v19 checkpoint. Developers can install the strands-decider package and use its command-line interface or integrate the model into Strands agents.
The second project, Strands Box, is available in developer preview. It provides an isolated environment for executing AI agents, combining operating-system-level isolation with Dogwood, an open-source policy language that defines which actions an agent can perform. The sandbox supports shell commands, Python execution, MCP tools, filesystem operations and network access, using kernel-enforced isolation to route agent interactions through a policy layer.
Analysis
Why This Matters
- Strands Decider 2B offers a low-latency, locally runnable option for routine agent decisions, an alternative to calling large general-purpose models for every step.
- Open release of weights, training data and scripts lets developers fine-tune and run the model on their own hardware, with confidence scores on each choice.
- Strands Box targets the safety problem of letting agents execute code and reach external systems, a known barrier to putting agents in production.
Background
AI agents are software systems that use language models to plan and act, for example running code or calling tools. Large models can be slow and costly when used for simple choices inside an agent loop, which has pushed interest in smaller, specialised decision models. Sandboxing, isolating untrusted code, is an established security technique. Strands combines a small decision model with a policy-enforced execution environment, part of a wider industry effort to make agent behaviour more predictable and safer to deploy.
Key Perspectives
AWS: Presents the two projects as complementary pieces of agent infrastructure, one for fast decisions between options and one for constrained, policy-checked execution, and highlights that both are open source.
Developers: Gain inspectable, local components they can modify, with published benchmark results to compare against. The confidence scores produced by Strands Decider 2B may help build more predictable agent behaviour.
Critics and skeptics: May ask whether 72.3% benchmark accuracy is high enough for production decisions, and note that Strands Box is still in developer preview. The reported latency figures come from specific hardware and small tasks, and may not hold across larger workloads.
What to Watch
- Whether Strands Box graduates from developer preview to general availability and which policy features ship with it.
- How Strands Decider 2B performs in real deployments compared with its reported JevBench accuracy.
- Uptake of Strands components and the Dogwood policy language in the wider agent development ecosystem.