Cloud Native Hyderabad

November Virtual Meetup - One Pod per Agent Doesn't Scale: Running AI Agents on Kubernetes with Agent Substrate

Capacity: 300
virtual
Free
Virtual event · Nov 19

Event date
Nov 19, 26
08:00 PM - 09:15 PM IST
Registration is open until Nov 19, 2026 at 7:30 PM IST.
Location
Virtual event
About this event

Kubernetes pods were built for long-running services, but AI agents sit idle most of the time and wake up in bursts. This talk shows why one pod per agent hits a wall, and how Agent Substrate, an open-source runtime, runs many agents on a small pool of pods.

Every team building AI agents runs into the same question: where does each agent actually run? On Kubernetes, the obvious answer is one pod per agent session. We tried that, and it topped out at about 20 sessions before we hit its limits.

The reason is simple. Pods were designed for long-running services. Agents, MCP servers and other bursty work sit idle most of the time, wake up for a few seconds, and need their state back when they do. That needs a different primitive underneath.

Swarup Donepudi will start from the basics: what a pod is good at, and why an agent is a different shape of workload. Then he will walk through Agent Substrate, an open-source runtime for agent sandboxes on Kubernetes: how it maps many agents onto a small pool of ready workers, suspends and resumes them with their state, and keeps each one isolated. You'll see it in action, and you'll leave knowing when one pod per agent is fine and when you need something like this.

Hosts
Organizers
Tags
LLM AI Kubernetes Platform Engineering