How to build and scale multi-agent AI systems on GKE

Google Cloud Tech · 2026-08-26

▶ Play the video

This is no longer on the current shelf — shelves rotate as new material clears the bar. The analysis below is unchanged. See what is featured now.

Why it earned a slot

Google’s GKE agentic system demo is a love letter to platform engineering, but it’s also a missed opportunity for broader accessibility. The ‘Kubernetes-sigs Agent Sandbox’ and ‘Model Context Protocol’ (MCP) feel like over-engineered solutions for a niche—distributed multimedia knowledge pipelines on GKE. The ‘Gemini 3.5 Flash’ part is slick, but why stop there? The real gold is the ‘diagnostic workflows as reusable skills’ idea, which is how we’ve been building agentic systems for years (just not at this scale). The Property Graph + GQL part is cool, but it feels like a vendor play for BigQuery—like they’re pushing their graph database instead of a general solution. Also, the ‘troubleshoot broken GKE deployments’ bit is useful, but it’s not clear how this generalizes beyond Google’s ecosystem. The ‘automated application evaluation’ claim is strong, but the demo feels like it’s solving for Google’s internal needs more than the general market.

The short version

Google’s GKE agent playbook is impressive—but is it too Google-y to matter outside their ecosystem?

Why it matters

This video is a case study in how platform engineering can scale agentic systems, but it’s also a reminder that ‘one-size-fits-all’ solutions rarely work. The ‘Agent Sandbox’ and ‘MCP’ are clever, but they’re built on Kubernetes—an ecosystem that’s already complex enough without adding AI-specific abstractions. The ‘diagnostic workflows as skills’ idea is brilliant (we’ve used similar patterns in our systems), but the execution feels tied to Google’s stack. Right now, the industry is split between ‘build your own’ (like we did) and ‘use the cloud provider’s tooling’—this demo leans hard into the latter. The ‘Property Graph’ part is a nice touch, but it’s not clear if this is a general solution or just a Google play. What’s missing is a discussion of how this handles *multi-cloud* or *hybrid* deployments—something we’ve struggled with in agentic systems.

My take

In agentic systems the biggest bottleneck isn’t the models but the *infrastructure*. Google’s approach here—treating diagnostics as ‘skills’ and embedding them in Kubernetes—is exactly what that gap calls for, and until now teams have had to bolt it together themselves. The ‘Gemini 3.5 Flash’ part is cool, but it’s not clear how this scales beyond Google’s ecosystem. The ‘MCP’ is a clever abstraction, but it’s not clear how it handles *latency* or *cost* at scale—two of the biggest pain points in agentic systems. What’s wild is how little this talk is about *agents*—it’s about *platforms*. That’s where the real innovation happens, but it’s also where most teams get stuck.

How it connects

• This connects to the ‘AI infrastructure’ push we’ve seen at events like NeurIPS 2025, where papers on *distributed LLM serving* were the most cited. Google’s approach here is a step toward ‘AI-native’ Kubernetes, but it’s not clear if this will become a standard or just another vendor lock-in.
• The ‘diagnostic workflows as skills’ idea mirrors what we’ve done in our systems, but the execution here is more ‘Google-centric’. This is a reminder that ‘agentic systems’ are still in the ‘build vs. buy’ phase—most teams are still figuring out what to build themselves vs. what to outsource.
• The ‘Property Graph’ part is a nice touch, but it’s not clear how this generalizes beyond Google’s ecosystem. This is a classic ‘vendor play’—Google pushing their graph database instead of a general solution. It’s a reminder that ‘AI infrastructure’ is still a moving target.

Bottom line

If you’re building agentic systems on GKE, this is a great starting point—but don’t expect it to work outside Google’s ecosystem. The ‘diagnostic workflows as skills’ idea is gold, but the execution is too tied to their stack.

Topics: multi-agent, gke, gemini, mcp