The software era of fixed-workflow SaaS is over; the harness era is here. If every company has access to the same AI models, the winners will be those who build the best systems for context retrieval, tool orchestration, state persistence, and observability. This is a pragmatic guide for engineers deciding where to invest their infrastructure effort.

Async RL has an expensive dirty secret — shipping the full model every step — but this post shows that 98-99% of bf16 weights don't actually change between steps. It lands a real TRL PR that encodes only the sparse delta as safetensors, uploads it to a Hugging Face Bucket, and lets vLLM fetch it. The result: payloads drop from 1.2 GB to 20-35 MB, and the trainer and inference server no longer need to share a network.