Salesforce post-trained a 120B Nemotron to take over its agents' multi-step reasoning

The training loop is the part of Salesforce's post an operator can use. A judge scored each simulated run on whether the customer's problem actually got resolved; when it did not, the model tried again, and each scenario ran thousands of times ^[11]. The scenarios covered 14 or more industries, including manufacturing, financial services, healthcare and travel, with synthetic customers given moods and personas, some cooperative and some frustrated ^[10]. Salesforce reports the outcome in words: the reinforcement learning improved multi-turn tool use, the ability to carry a task across many steps, tools and turns of conversation ^[12]. There is no benchmark score, resolution rate, latency figure or cost per task in the post ^[17].

This has been coming in pieces. For most of Agentforce's history one general model did the reasoning for almost every task, including intent detection, prompt-injection screening, toxicity, evaluation and reranking ^[5]. Over eighteen months Salesforce moved those narrower jobs onto a series of smaller models each built for a specific job, and left the core multi-step reasoning where it was ^[6]. The multi-step reasoning is the last piece to move ^[18].

The argument in the post is about variance. A general model works each task out from scratch every time it gets one and, because it reasons probabilistically, may not do it the same way twice ^[19]. Salesforce wrote that a frontier model "reasons about your business from the outside, applying broad intelligence to work it has never actually done" ^[15], and that "governed, repeatable work favors a model that has learned the exact work to be done, and always leverages that expertise" ^[16]. Clearing a refund and qualifying a lead are the examples it reaches for ^[20].

Nemotron 3 Super is a 120-billion-parameter open model ^[3]. Salesforce says the open weights gave it two things a closed frontier model would not: direct access to post-train and adapt the model, and NVIDIA's published record of the datasets the foundation was trained on ^[7]. The company describes what it got as "an open, US-trained model we control and can inspect" ^[8], and says it drew on 27 years of running the same workflows it wanted the model to learn ^[13].

For the person who has to answer for a swapped-out reasoning layer, the scenario worth reading is the one Salesforce singles out. It deliberately put the model in simulated situations where the right tool was not available, and calls that the case enterprises worry about most ^[14].

Two properties sort a portfolio of agent processes: whether the work has to run the same way every time, and whether it resembles the workflows Salesforce rehearsed in simulation. Where both hold, Koa is the better fit, and the check is a team's own transcripts scored the way the judge scored the simulations, on whether the problem got resolved. Where the second does not hold, the reasoning was rehearsed on someone else's work, and the general model should stay in place until the transcripts say otherwise. Salesforce says customers can now use Koa to run their AI agents' processes ^[4].