Fleet as Post-Training Operating Layer

Control is moving from who can afford the biggest pretraining run to who can shape model behavior after the base model exists. In practice, that means the scarce asset is not just compute, it is the system for generating hard tasks, scoring outputs, comparing versions, and deciding whether a model is ready for production. That is why Fleet can grow from hosting environments into the operating layer where labs tune, test, and ship agents.

• Fleet is already extending in that direction. Harbor supports arbitrary agents, shared benchmarks, large batches of experiments, and RL rollout generation, which are the core mechanics of post-training rather than simple environment hosting.
  1 sacra 2 sacra
• The broader market is reorganizing around this workflow. Human data vendors like Prolific, Micro1, and Surge are moving from basic labeling into evaluation, red-teaming, and specialist feedback, because model quality now depends on better reward signals and expert review loops.
  3 sacra 4 sacra 7 sacra 8 sacra
• Major labs are reinforcing the same pattern. OpenAI publishes reinforcement fine-tuning workflows built around graders and eval pipelines, and Anthropic has made model evaluations central to both alignment work and release decisions. The bottleneck is increasingly measurement, not just model size.
  5 openai 6 anthropic 9 openai

This pushes the market toward full post-training stacks. The winning platforms will combine task environments, synthetic and human generated datasets, reward models, regression testing, red-teaming, and deployment gates in one loop. If Fleet keeps building across that loop, it can become infrastructure that customers run every time a model is updated, not just when an environment is first created.