Salesforce introduces Koa Decisions for millisecond agent routing

Salesforce introduces Koa Decisions for single-token agent routing 📅 November 6, 2026 • Salesforce What was introduced Salesforce has introduced Koa Decisions, a model specialised for routing requests inside agent systems, together with the next version of HyperClassifier. The task is deliberately narrow. The model reads a request together with a list of categories, returns a decision in a single step, with the answer often being a single token. A classifier that only has to pick one label does not need to generate a sentence, and removing that generation step changes what a routing layer costs in time and in compute. Measured speed Salesforce measured Agentforce routing at about 26 milliseconds, against about 1.4 seconds for a general purpose language model. On those two figures that works out to roughly 50 times faster, a multiple Salesforce does not itself publish. The comparison reflects a difference in the shape of the task rather than a larger model. A general purpose model answers with generated text and reasoning, while a decision model is trained to emit one answer from a closed set of options. Training and input types The model was built by fine-tuning the open-weight Gemma model from Google with new training techniques. It also adds multimodal support, so images and audio can take part in the routing decision. That matters in service scenarios where a request often arrives as a photograph, a screenshot of an error or a short voice message, and a text only classifier would lose that part of the input before the decision is made. Rollout and research Koa Decisions is the new version of HyperClassifier that routes requests inside Agentforce, and Salesforce plans to extend it beyond routing, into flows, Apex and prompt templates, plus an API for developers. It joins the Koa model family, which includes a reasoning model built with NVIDIA. The research paper describing the approach is published on arXiv under the number 2609.36115. What it means in practice The practical reading is that agent platforms are starting to divide model work by task type. A large general purpose model stays in the loop for open ended generation, while faster and cheaper specialised models take over routing, extraction and classification. For platform teams this means the routing layer stops being the slowest part of the path and becomes a component that can be measured, tuned and priced on its own. Open questions remain about how stable the decision stays when the list of categories keeps changing and how much of the multimodal support is exposed in the first release. The single token output also makes the result easy to log and easy to pass to the next step, which is useful when a routing decision has to be explained later. Teams should still compare the specialised model against a general purpose model on their own category sets before moving traffic. IMPACT Moving routing away from general purpose generation is a practical way to cut latency in agent platforms, and Koa Decisions shows how a specialised decision model can be measured and priced on its own. Sources: https://www.salesforce.com/news/stories/koa-decisions-salesforce-decision-models/ — https://arxiv.org/abs/2609.36115 — https://unite.ai/salesforce-introduces-koa-decisions-with-image-and-audio-support Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky