AWS Releases New Open Source AI Agent Tools

The release includes a specialized decision-making model and a sandboxed execution environment for agent security.

Updated on Oct. 16, 2026 in Artificial Intelligence

Isometric editorial illustration of floating modular cubes in deep teal and mustard, representing secure AI sandbox architecture. — AWS released two open-source AI projects, Strands Decider 2B and Strands Box, to enhance agent security and reliability through structured decision-making models. AI Illustration. Upload story photo >

Live Poll

Do you trust automated sandbox tools to effectively control the actions of AI agents?

AWS has released two open-source projects, Strands Decider 2B and Strands Box, to improve AI agent decision-making and security. Strands Decider 2B is a two-billion-parameter model available with its training data, while Strands Box is a sandbox environment currently in developer preview.

Why it matters

These tools address the need for reliable execution boundaries and efficient decision-making in AI agent workflows. By shifting from generative text to structured selection, developers can reduce reliance on prompt engineering for agent reliability.

Strands Decider 2B is a two-billion-parameter model based on Qwen3.5-2B, utilizing a rank-16 LoRA adapter to score candidate answers rather than generating text. It demonstrates a median decision latency of 115 milliseconds on an NVIDIA RTX 3090 and 153 milliseconds on an Apple M3 MacBook.

The players

AWS

A cloud infrastructure provider offering compute, storage, and AI model hosting services.

The details

Strands Decider 2B replaces the language-model head of Qwen3.5-2B with a pointer head to score predefined options. Strands Box provides an execution sandbox that routes agent interactions through a policy layer using the Dogwood policy language to enforce kernel-level isolation. This setup supports Python execution, filesystem operations, and network access while restricting actions based on execution history.

October 16, 2026: AWS released Strands Decider 2B and Strands Box.

The Tech Race

The release marks a shift toward specialized models designed to outperform general-purpose agents on JevBench metrics. AWS is positioning these tools to compete with custom agent-workflow frameworks that prioritize low latency and security isolation over broader creative capabilities.

Developers building autonomous agents can now integrate the open-source Strands Decider 2B model to handle logic tasks with sub-200 millisecond latency on local hardware. Strands Box is available immediately as a developer preview for those requiring sandboxed execution environments for Python or shell-based agent tools.

The takeaway

The move suggests a focus on moving AI agents from unreliable text-based instruction followers to secure, logic-focused task executors. Watch for future JevBench leaderboard updates to see if the two-billion-parameter size remains a competitive threshold for agent-specific decision models.