Loading page…
Strands Decider 2B can run locally and returns confidence scores alongside predefined choices. AWS’s pitch is faster, potentially cheaper workflow decisions—but preserving intelligence remains the challenge.
Listen to this story
AWS is packaging a narrower agent component for workflows where the next action can be selected from a fixed menu rather than generated as open-ended text. Strands Decider 2B is open source and small enough for local use; AWS presents it as a way to reduce latency and potentially cost, while preserving enough language understanding to make reliable choices remains an engineering tradeoff. The release turns Marc Brooker’s experiment into an AWS offering as decision-focused models emerge alongside general-purpose LLMs.
The model returns a choice from predefined options with a confidence score, rather than an explanation.
Brooker’s prototype briefly topped the Jevbench ranking among models of its size, according to TechCrunch.
Brooker says the challenge is improving accuracy and confidence calibration without losing language understanding and knowledge.
An AI agent does not have to write an answer every time it decides what to do next. Amazon Web Services has released Strands Decider 2B, an open-source model that chooses among predefined options and returns confidence scores. Available now and small enough to run locally, it targets workflow steps AWS says can be handled without the capability or expense of a full language model.
The project began with AWS distinguished engineer Marc Brooker seeing TypeSafe’s Jev and trying to build his own version. Brooker’s prototype briefly reached the top of the Jevbench ranking among models of its size, according to TechCrunch.
Amazon engineers then cleaned up the experiment and released it through Strands Labs, which develops tools and protocols for deploying AI agents.
Brooker told TechCrunch that conversations with AWS customers revealed the need for this kind of component. Their agent workflows did not always require a full-featured large language model, or LLM—the kind of model used for open-ended text generation. The narrower question was what an agent should do next, given where it already was in a workflow.
Its output is therefore a decision rather than an explanation.
Brooker described a careful balance between improving accuracy and calibration—how well confidence tracks correctness—and preserving language understanding and knowledge. Those broader abilities are what make the model general-purpose and useful, he told TechCrunch. Restricting the available answers does not remove the need to understand the information behind a choice; the engineering goal is to improve decision performance without losing that foundation.
TypeSafe founder and CEO Diogo Almeida offered a skeptical response to the emerging competitors, telling TechCrunch that people might underestimate the difficulty of making these models smart.
Loading discussion...
Join the conversation
Explain when a confidence score would—or wouldn’t—be enough.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.