AWS Releases Open-Source Decision Model for AI Agents
Amazon Web Services has released Strands Decider 2B, an open-source decision model inspired by TypeSafe's Jev, according to TechCrunch. The report, published Oct. 1, 2026, places the AWS release in the same week OpenAI announced a similar offering. Decider does not generate open-ended text. It sorts between pre-decided options and provides a measure of confidence in its choice.
Strands Decider 2B is fully open-sourced and available now. It is small enough to run locally, on your own hardware rather than over the network. It is built on the torso of the Qen3.5-2B LLM, meaning it reuses the base structure of that 2-billion-parameter model, and delivers calibrated choices instead of generating text. Calibrated means the confidence score lines up with how often the choice is correct. That separation is the point. Generation proposes language. Decision selects among bounded alternatives with an explicit confidence signal.
The project started as a homebrew effort from Amazon distinguished engineer Marc Brooker. After seeing Jev, Brooker built his own take on such a model. His initial version briefly reached the top spot on the Jevbench ranking for models of its size. Amazon engineers then cleaned up the project for release as an offering from Strands Labs. Strands Labs is an organization developing new tools and protocols for deploying AI agents, software systems that can plan and carry out multi-step tasks.
The lineage traces to TypeSafe. The company named its model Jev after economist William Stanley Jevons. TypeSafe's CEO and founder is Diogo Almeida.
The broader context here is the split between generation and selection inside agentic stacks, the layered software that lets agents plan and act. Practitioners already constrain agents with tool schemas, policy routers, allowlists, and retrieval boundaries, essentially rulebooks that define which tools an agent can call and what data it can pull in. A dedicated decision model formalizes that constraint. Instead of parsing free text and hoping the choice is consistent, the agent delegates the choice to a component trained to output a distribution over known options, a spread of probabilities across the allowed picks. For engineers working with orchestration layers, function calling, and multi-step planning, calibration becomes operational infrastructure. A well-calibrated confidence score can gate execution, trigger human review, or select fallback logic.
In my view, local execution and open source change the iteration loop. A small decision model that runs locally can sit adjacent to a larger generator without adding a network hop for every branch point. It can be evaluated, versioned, and swapped independently. It invites direct measurement of choice quality under distribution shift, prompt variation, and tool-set growth, or how it holds up when data, instructions, or available tools change. The long arc in systems work favors components with narrow contracts that fail legibly. Generators remain powerful for summarization, drafting, and open-ended reasoning. Deciders handle the moment where the system must commit. Watching that division of labor become explicit, testable, and openly available gives teams more leverage over reliability than a larger context window, the amount of text a model can consider at once, alone.
