Amazon Web Services has released Strands Decider 2B, a decision model that selects among predefined options and reports its confidence in the choice. Designed around decisions rather than text generation, it offers a way to handle the next-step choices within an AI agent’s workflow.
Choosing a next step instead of generating text
Strands Decider uses part of the Qen3.5-2B language model as its foundation, but its output is a calibrated choice rather than generated prose. The available answers are defined in advance, and the model supplies a measure of confidence alongside its selection.
Amazon distinguished engineer Marc Brooker said conversations with AWS customers revealed a need for this approach: their agentic workflows did not require the capabilities of a full-featured language model at every step. He described the decision model as a way to choose what should happen next based on a workflow’s current state.
Brooker said the bounded set of answers and confidence scores could make those workflow decisions more reliable, with lower latency. He also identified a development challenge: improving decision accuracy and calibration while preserving the model’s understanding of different languages and the knowledge that supports broader use.
From a Jev-inspired experiment to Strands Labs
The project began after Brooker encountered TypeSafe’s Jev and tried building his own decision model. Amazon engineers subsequently refined the experiment for release through Strands Labs, which develops tools and protocols for deploying AI agents.
TechCrunch reported that the original experimental project briefly held the top position on the Jevbench ranking among models of its size. That result concerned the early project and its size category, rather than a comparison placing the released model ahead of all other models.
Available now for local use
Strands Decider 2B is fully open-source and already available. It is small enough to run locally. Brooker said the decision-model approach could also potentially reduce the cost of workflow steps that do not need a full-featured language model.
