Scale AI pivot to agent simulators

This shows Scale is moving from selling labeled examples to selling practice worlds for agents. In the old workflow, humans tagged an image or ranked a text answer. In the new workflow, the model has to click through software, use tools, recover from mistakes, and get scored on whether it finished the job. If nearly half of new projects now use RL environments, customer demand is shifting toward training models to do work, not just predict tokens.1 2 3 4

• An RL environment is closer to a flight simulator than a spreadsheet of labels. Scale’s product focus now includes simulated tool use and computer use workflows, where labs can watch an agent take actions step by step, grade outcomes, and use those traces for post training and evaluation.1 4
• This also changes the competitive set. Scale is no longer competing only with classic labeling vendors like Appen or CloudFactory. It now overlaps with newer agent training companies like AfterQuery, plus human feedback specialists like Invisible and Mercor that sell expert judgment for reasoning and post training workloads.1 2 5 6
• The money and labor model shifts as well. Raw labeling is paid per annotation, but RL style work needs richer assets, domain experts, and reusable tasks that mimic real software jobs. That supports higher value contracts because customers are buying a test bed for agent behavior, not just a bucket of tagged data.1 3 4

The next step is a broader market move from human labels toward packaged training environments, evaluations, and agent benchmarks. The winners will be the companies that own realistic task worlds and the worker networks needed to score them, because that infrastructure sits directly in the loop of making AI agents reliable enough for production use.1 3 4 5