AWS open-sources Strands Decider 2B model and Strands Box sandbox for AI agents

AWS has published two open-source projects aimed at the two hardest parts of building an AI agent: deciding what to do next, and keeping the execution of that decision under control. The first, Strands Decider 2B, is a small model trained to make routing and planning choices inside an agent workflow. The second, Strands Box, is a sandbox that governs how an agent runs code and reaches outside systems. InfoQ's Daniel Dominguez reported the release.

The split reflects a problem teams hit once agents move past chat. A general-purpose model can call a tool, but the orchestration layer still has to decide which tool, in what order, and whether to stop. Strands Decider 2B is pitched at that narrow job rather than general reasoning, which matters because a 2-billion-parameter model can run on far less hardware and with lower latency than a frontier model invoked for every branch point. The name suggests it slots into the Strands agent framework AWS has been building out.

Strands Box addresses the other half. Agents that execute generated code or touch external services need a boundary around what they can reach and what they can run. A sandbox is the standard answer, but it is often bolted on late and inconsistently. Shipping one as a standalone open-source component gives developers a defined execution surface to test against, rather than each team rolling its own isolation.

The two projects are separable, and that is the practical point. A developer can adopt the decider model without the sandbox or the sandbox without the model, which lowers the cost of trying either. For AWS, open-sourcing both is a bet that agent builders will standardize on its primitives and, eventually, its hosted services. The company has been steadily assembling agent tooling around Strands, and these pieces fill gaps that previously pushed teams toward third-party frameworks.

Small decision models are a growing category. Running a dedicated few-billion-parameter model for structured choices can cut cost and latency compared with calling a large model at every step, and it keeps sensitive routing logic closer to the application. Whether a 2B model is accurate enough for complex multi-step workflows is the open question; the release gives teams a way to measure that on their own tasks instead of taking a vendor's word for it.

What to watch next is adoption and benchmarks. AWS has not yet detailed how Strands Decider 2B performs against general-purpose models on tool-selection tasks, and the sandbox's isolation guarantees will face scrutiny from teams running untrusted generated code. Both are open source, so the answer will come from developers testing them in production rather than from a launch post.