Agent infrastructure is getting its own stack
Rivet (YC W23) launched an orchestration layer that runs agents on S3, arguing that sandboxes are too slow to boot just to recover agent state. One commenter said they already moved a Cloudflare Durable Object workload onto it inside their own VPC and had a pleasant time. Nvidia published Nvidia Kubernetes Engine, and the top reaction was whether regular Kubernetes can't already schedule GPUs, plus a complaint that the page reads like an AI-generated data dump. Mbodi AI (YC P25) is hiring research and robotics engineers.
The pattern here is plumbing. Agents need durable state, cheap pause and resume, and GPU scheduling, and every vendor from startups to Nvidia is trying to own a layer of that. Also visible: Unleash, an 11-year-old open-source feature flag server, got a comment that 'configuration as a service' companies must be struggling now, while another noted LaunchDarkly is a $3B company.
The weak spot in the discussion is differentiation. Commenters keep asking what the new layer does that existing tools don't.
So what?
If you sell infrastructure for agents, lead with a concrete pain, like state recovery cost, not a platform story. Hype fatigue around Kubernetes-flavored announcements is real, and 'can't I already do this?' is the question you must answer in the first paragraph.
Read these
Launch HN: Rivet (YC W23) – Scale and Orchestrate Agents on S3
Nvidia Kubernetes Engine (NVKE)
Mbodi AI (YC P25) Is Hiring Research and Robotic Engineers
We've built an open-source feature flag server for 11 years