Google: Size GKE Node Pools for Agent Sessions, Not Requests
Start here
Google published a technical guide arguing that AI agent sessions on Google Kubernetes Engine (GKE) should be sized differently from web requests, because agents spend most of their time waiting rather than computing. Google cites a GitHub Copilot trace study finding that agents wait for the user about 80% of the time. The guide centers on Agent Substrate 1.0, which Google says is now generally available on GKE. It suspends idle actors to object storage and frees their workers, letting more agents share the same compute. According to Google, Agent Substrate does not suspend an idle actor on its own: the agent harness, or the service in front of it, must call the suspend API after an idle timeout. If idle actors hold every worker, a request for another actor waits up to 5 seconds by default and then fails with HTTP 503, per Agent Substrate documentation. Google also describes AX, a companion component it says is still in active development ahead of its first stable release, and invites feedback and issues on GitHub.
Generated from reports · updated 1 h ago
Report timeline
Follow the reports to see the event from different sides.
- Google Developers BlogGoogle: Size GKE Node Pools for AI Agent Sessions, Not Requests
Google argues AI agent sessions should be sized differently from web requests on GKE, since a GitHub Copilot trace study found agents wait for the user about 80% of the time. Agent Substrate 1.0, now generally available on GKE, suspends idle actors to object storage and frees their workers, letting more agents share the same compute.
Heat for this event
Not enough continuous observations yet to draw a trend.