SALT LAKE CITY – Google’s recent donation of its Agent Substrate project to the Cloud Native Computing Foundation (CNCF) was somewhat overshadowed by some Nvidia news (no, not that news!), but attendees at this week’s KubeCon+CloudNativeCon event in Salt Lake City lumped both moves as being key toward securing AI infrastructure.
Agent substrate was designed to decouple stateful agent sessions and identities from active compute, taking a snapshot of idle agents for storage and then being able to restore those agents in less than 500 milliseconds. This is targeted at managing the idle state of those agents so that they do not burn expensive compute resources, a challenge in a Kubernetes-managed environment that might be tasked with managing thousands of agent-infused pods.
“Think of it as a hotel room rather than a house: you don't keep a room when you leave town, and a new room is ready when you come back,” Drew Bradstock, VP of product management, orchestration, and Kubernetes at Google, explained in a blog post.
Google donated the platform’s base configuration into the CNCF “Sandbox” program, which is the entry point for projects at that open-source organization. However, Agent Substrate does have a bit of work behind it having been released by Google into public development in May, a move that has garnered interest from more than 100 contributors, has seen the project also injected into the CNCF kagent AI programming framework project, and for Google to have unveiled Agent Substrate 1.0 as part of its managed Kubernetes Engine (GKE) for public use.
CNCF Executive Director Jonathan Bryce during a press briefing at this week’s KubeCon+CloudNativeCon event noted that Agent Substrate is significant in that it helps counter the agentic AI efficiency lull that has happened over the past year.
“A lot of folks when OpenClaw first came out at the end of last year and the beginning of this year, we were buying Mac Minis … Mac Minis were sold out all over the world, and people were running them on a Mac Mini,” Bryce said of OpenClaw’s management of AI agents. “But that agent, even though it's long running, even though it may do work arbitrarily day or night, it's not necessarily using that Mac Mini at 100% of the capacity over that time. So you end up with sort of a lot of resources that you're not maximizing.”
A package changing open source #
More importantly, Agent Substrate can be partnered with OpenShell to provide a compelling cloud-native package for running and securing AI agents.
“These are projects that are directly oriented around taking these powerful and non-deterministic compute agents and figuring out how do we contain them and how do we go from a point of zero trust at first but being able to give them the access they need to make progress, and that's a really hard challenge,” Bryce said.
Bryce added that this challenge was exemplified by a recent CNCF survey that found while coding agents were “extremely widely adopted, … if you look at who's got an agent in their personal life that's doing anything significant [it’s a] much smaller amount of the audience, and that's because: is it safe? Can I trust it? Do I really want to give it the keys to anything? And for most people … the answer is definitely no.”
“What's great about these projects, though, is if you look at something like OpenShell and Agent Substrate, they sort of just snap into existing cloud-native infrastructure,” Bryce continued. “And with 82% of enterprises running Kubernetes right now, and things like Open Telemetry and the set of commonly deployed cloud-native projects, that gives you a huge advantage if you are one of the enterprises who's adopted that stack, because now you can plug OpenShell in and start experimenting with agents in a way where it's using the same security policies, network policies, telemetry that you've already set up and that you know is well baked and well defined. Same thing for Agent Substrate.”
The projects also highlight the rapidly changing evolution timeline for open-source projects.
“These are projects that are very early for sure, but moving super quick, and I think give us a glimpse of where we're going to see AI infrastructure developing over the next six to 12 months,” Bryce said. “One of the things that I'm most excited to see is how people start to put these together and we get a stack that kind of gels. Because that's when we, I think, historically in open source, have seen the greatest advancement and the greatest adoption is when we get those patterns that work.”