Amazon Web Services has released an open-source decision model that gives AI agents a faster, cheaper way to choose between predefined options without calling a full-scale language model at every step.
Strands Decider 2B is a roughly 2-billion-parameter model that returns structured choices along with confidence scores. It is small enough to run locally and arrives as developers show growing interest in decision-focused models for software automation.
Built for agent workflow decisions
Amazon distinguished engineer Marc Brooker began experimenting with the approach after seeing TypeSafe AI’s Jev model. He later described the project’s engineering process in a technical post about building a System One-style model.
Brooker said conversations with AWS customers showed that many agent workflows do not need the full capabilities, latency or cost of a frontier LLM at every step. A smaller decision model can instead answer questions such as what an agent should do next within a closed set of possible actions.
The model is built on Qwen3.5-2B but is optimised to make calibrated choices rather than generate open-ended text. During development, it briefly reached the top position among similarly sized models on a Jev-focused benchmark ranking.
Decision models are becoming a broader category
Strands Labs released Strands Decider, Amazon’s open-source effort focused on tools and protocols for building and deploying AI agents.
The release comes as more researchers and companies experiment with decision models inspired by TypeSafe’s Jev. These systems aim to preserve enough language understanding and general knowledge to make useful decisions while reducing latency and cost for repetitive workflow steps.
Brooker said the challenge is balancing speed and calibration with broader intelligence, including language understanding and general knowledge. TypeSafe CEO Diogo Almeida has similarly argued that making a decision model fast and inexpensive is easier than making its outputs consistently intelligent.
With Strands Decider 2B now available as an open-source model, developers have another option for separating routine agent decisions from tasks that still require a larger reasoning model.
