← All AI news
GoogleAgents

Long-lived agent sessions need a different capacity plan than web requests

Google’s GKE cost model compares dedicated pods with parking idle agent sandboxes and restoring them on demand.

THE UPDATE

What changed

Google explains how Agent Substrate and AX can snapshot an idle agent sandbox to object storage, release its worker and restore the session later. In the article’s model for 100 coding agents and two prompts per hour, parking reduced the node bill by 40 percent.

One useful next step

Measure one week of prompts per hour and session idle time. Compare dedicated pods and parked sessions at the same CPU guarantee, including the busiest hour, restore behavior and the cost of a control plane before changing infrastructure.

Define workload measures before a platform trial ↗

Before you dive in

The 40 percent figure comes from Google’s stated assumptions and list-price model. The article says parking stops paying in its model around 4.5 prompts per hour, and AX is still in active development ahead of a first stable release.

GO TO THE ORIGINAL

Source & announcement date

Google ·

Google: Agents are sessions, not requests — size GKE node pools for AI agent sessions ↗ (opens in a new tab)

This brief summarizes the linked announcement. “The AI Superpower take” and the next step are our editorial interpretation, not claims from the source.

KEEP YOUR CURIOSITY GOING

A few more worth knowing.

All AI news ↗