Salesforce introduces Koa Decisions, a Gemma-based decision model built to route Agentforce requests in milliseconds

On November 6, 2026, Salesforce introduced Koa Decisions, the next version of HyperClassifier, the decision model that sends each request inside Agentforce to the part of an agent built to handle it. Given a request and a list of categories, it returns one choice with a confidence score in a single step instead of writing an answer token by token. Salesforce's data science team built it by applying new training techniques to Google's open-weight Gemma model, and it adds support for image and audio input.

Salesforce says routing a request in Agentforce took about 26 milliseconds with its decision model, compared with about 1.4 seconds for a general-purpose LLM, and that new categories work without retraining. It plans to extend Koa Decisions beyond routing to Flows, Apex and prompt templates, plus an API for developers, but gave no dates or pricing. You can try it in the Koa Decisions demo on Agentforce Labs with a free Trailhead account, and a research paper describes the method. The model joins the Koa family, which also includes a separate reasoning model built with NVIDIA.