AWS Introduces Strands Decider 2B and Strands Box for AI Agent Development

AWS has released two open-source projects aimed at different parts of AI agent development: Strands Decider 2B, a small model that makes discrete decisions inside agent workflows, and Strands Box, a sandbox for constraining what an agent is allowed to do once it starts acting. InfoQ reports the announcement via Daniel Dominguez. The two projects target agents that call tools, execute code and reach out to external systems.
Strands Decider 2B is a two-billion-parameter model built to choose among predefined options rather than emit long generated text. Its intended jobs are model routing, tool selection, classification, guardrails and context management. The interaction pattern is unusual for an LLM: instead of asking a general-purpose model for a free-form answer, the developer supplies a candidate set and receives a decision plus confidence scores. Architecturally it derives from Qwen3.5-2B, with the language-model head swapped for a pointer head that scores candidate answers, and it is fine-tuned through a rank-16 LoRA adapter. AWS has published the weights, training data and scripts, so the model can be run and modified locally.
AWS reports a median decision latency of roughly 115 milliseconds on an NVIDIA RTX 3090, and 153 milliseconds for small tasks on an Apple M3 MacBook. The company evaluated the model with JevBench, reporting 72.3% accuracy across 231 public tasks for its v19 checkpoint. Developers can install the strands-decider package and drive it through a command-line interface or wire it into Strands agents.
A pointer head, in this context, is a lightweight output layer that assigns scores to a supplied list of candidates rather than generating tokens one at a time. That is much cheaper than asking a large model to reason in prose, and it fits agent loops where the decision is narrow, such as which tool to call or whether a request should be blocked. In practice, this suggests Decider 2B is meant to sit alongside a larger model: the big model plans or writes, while the small one makes the frequent, low-stakes routing calls. A likely trade-off is that it only answers the questions you framed as options, so anything outside that set falls to other components.
Strands Box, the second project, is in developer preview and provides an isolated environment in which agents run. It pairs operating-system-level isolation with Dogwood, an open-source policy language that spells out which actions an agent may take. The sandbox covers shell commands, Python execution, MCP tools, filesystem operations and network access. Kernel-enforced isolation routes agent interactions through a policy layer, and rules can allow or deny actions based on their content and on execution history. AWS gives the example of a policy requiring an agent to pass tests before it pushes code to a repository, or one that limits how often external APIs are called. Because these controls live outside the language model, they reduce reliance on prompt instructions to hold execution boundaries.
AWS says Strands Box works with different models and agent frameworks, locally or in the cloud, so teams can add execution policies without replacing their existing agent architecture. Compared with container-based sandboxes, Firecracker microVMs or hosted execution environments, the distinguishing feature is that policies can reason about tool arguments and prior actions, not just filesystem, process and network reach.
Decider 2B's closest counterpart is Jev from TypeSafe AI, another specialized model for structured decisions rather than long responses. Both cover classification, routing and tool selection, and both return confidence scores. The split is deployment: Decider 2B is open source and runs locally, while Jev is reached through a managed API. AWS used JevBench to evaluate Decider, but per InfoQ the published results do not establish a direct performance advantage over Jev.
Early developer feedback has centered on the practical friction of adding a small decision model to an agent stack. In a hands-on tutorial, AWS Community Builder Yuuki Yamashita found that Strands Decider 2B could stop an agent from making an unsupported weather-tool call. GitHub discussions have flagged issues around concurrent requests, Apple Silicon compatibility and cold-start latency. AI Builder Dipen Vaja commented on X that he would prefer an agent hit a wall mid-run than finish a bad path, describing Strands Box as a kernel sandbox that routes shell, Python, tools and network through one policy layer, allowing a rule to block outbound HTTP after an agent reads a customer file.
Taken together, the projects push the Strands ecosystem past orchestration into two adjacent concerns: choosing the next action, and policing how that action is carried out.
Why it matters: teams building tool-using agents get two separable pieces they can adopt independently, a locally runnable decision model for routing and guardrail calls, and an execution layer that enforces rules the model cannot talk its way around. That matters most for agents that touch shell, filesystems or network egress, where prompt-level instructions are weak protection. The open questions are practical rather than conceptual: whether cold-start and concurrency issues reported in early testing are resolved before production use, and how a two-billion-parameter local model compares on real routing tasks against a managed API alternative.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
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 projects address different aspects of building agents that use tools, execute code, and interact with external systems. By Daniel Dominguez