AWS has released two open-source tools aimed at making AI agents easier to steer and safer to run: a small model that chooses between options, and a sandbox that enforces rules on what an agent can do. Together, they tackle two practical headaches in agent development: choosing an action and controlling its execution.
AWS AI Watch analysis
What happened
Strands Decider 2B is a two-billion-parameter model for tasks such as tool selection, routing and classification. Developers provide a set of choices; the model returns a decision with confidence scores rather than generating a long response. News.LAVX reports that AWS has released the model’s weights, training data and scripts, allowing developers to run or modify it locally. The outlet cites AWS-reported results of 72.3% accuracy on 231 JevBench tasks and median decision latencies of about 115 milliseconds on an RTX 3090 and 153 milliseconds on an Apple M3 for small tasks.
Strands Box, in developer preview, provides isolated execution for agents using shell commands, Python, MCP tools, files or network access. Its Dogwood policy language lets developers set rules around actions and execution history. For example, a policy could require tests to pass before an agent pushes code, or limit how often it calls an external API. News.LAVX attributes the release details to InfoQ.
Why it matters
A model’s instruction to behave is not the same as an enforceable rule. Strands Box puts controls outside the model, where developers can restrict actions even when an agent is running different models or frameworks. Decider 2B addresses a separate need: making a bounded choice quickly, without asking a general-purpose language model to write an essay before picking a tool.
The combination gives developers more control over both decisions and execution, while the open-source model can be run locally. But the reported benchmark does not establish that Decider outperforms alternatives, and Strands Box is still in preview. Useful ingredients, then, not a finished safety guarantee in a very tidy box.
Our read
This is a substantial addition to the agent toolbox: one component helps choose the next move, and another can constrain what happens next. Developers building agents should look closely at the policy layer, while treating benchmark figures as a starting point rather than a verdict. The real test is whether the controls hold up across ordinary deployments, awkward edge cases and concurrent requests.
What to watch
- Whether Strands Box moves beyond developer preview and which environments it supports.
- How developers assess its policy enforcement, compatibility and cold-start behaviour.
- Whether Decider 2B’s benchmark results translate to real agent workflows.
- How the tools compare with managed decision APIs and existing sandbox approaches.
Discussion spark: Should agent developers rely on a model to choose safe actions, or require an external policy layer to block actions regardless of what the model decides?
Sources and evidence
- AWS open-sources Strands Decider 2B and Strands Box for AI agents – news.lavx.hu (16 October 2026, 16:39 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)