2026 Will Be The Year Of Agentic Workloads In Production On Amazon Eks

When AWS launched Elastic Kubernetes Service successful 2018, nan target assemblage was early adopters who wanted a managed power level and to grip everything other themselves. Over nan past 8 years, that has changed rather a bit.

“We’re starting to get into much nan precocious majority, moreover nan laggards,” said Mike Stefaniak, elder head of merchandise guidance for EKS and ECR astatine AWS. “People travel into Kubernetes. They don’t person monolithic level teams. They don’t want to negociate each azygous point themselves.”

In this section of The New Stack Makers, I sat down pinch Stefaniak to talk really AI workloads are reshaping Kubernetes, why AWS open-sourced a Model Context Protocol (MCP) server for EKS, and what’s really happening pinch agentic AI successful accumulation environments.

Tool Names Matter

Earlier this year, AWS Labs released an MCP server for EKS arsenic a Labs project.

“Some of nan early feedback that I recovered absorbing is that nan existent instrumentality names that you’re putting successful your MCP server matter rather a bit, wherever you don’t conscionable needfully want to transcript each API telephone that EKS has, because nan LLM [large connection model] tin fig that retired itself,” Stefaniak explained. “Troubleshooting, runbooks, really to deploy CloudFormation stacks to really tally a afloat exertion — that’s overmuch much absorbing that we put successful our MCP server than conscionable create cluster, deploy pod. We’ve evolved nan instrumentality names we launched pinch rather a bit, and now we’re moving connected managing it to make it much enterprise-ready.”

The squad has besides launched a hosted knowledge guidelines that contains years of support cases and soul logs that it tin now provender to an supplier that whitethorn beryllium trying to lick a akin problem. “We’ve seen each imaginable measurement that a node tin fail, that a cluster tin fail. If we tin bundle each that into an supplier and our MCP server, customers tin lick problems without having to unfastened a support ticket,” Stefaniak said.

Still Experimental, but Getting Real

While “agentic AI” continues to predominate galore convention — and particularly keynote — conversations, Stefaniak argues that astatine slightest for his customers, it’s still very early connected successful this game. Those customers are still very overmuch focused connected regular LLM inferencing connected EKS.

“I would opportunity agentic AI is much nan frontier,” he noted. “Most users coming still person a quality successful nan loop. Troubleshooting cases, for illustration — fto nan supplier fig it out, springiness a suggestion. If you’re going truthful acold arsenic to conscionable fto nan supplier effort to hole it itself, that’s further than astir group are.” 

That’s particularly existent for customer-facing applications, he said. “The soul platforms, they’re much consenting to beryllium bleeding separator because it’s internal. For customer-facing applications, there’s a batch much be aware there.”

What’s Next?

But Stefaniak does judge that this will alteration soon. “2026 will beryllium nan twelvemonth of accumulation deployments for agentic workloads, whereas 2025 was much accepted LLM inferencing and experimenting pinch agentic workloads,” he said.

For teams that want to research pinch agentic workloads, he recommends AWS’s unfastened root Strands SDK for penning agents successful Python, starting pinch an outer exemplary endpoint. 

His favourite “Hello world” workout for getting started pinch agents? Building a Kubernetes troubleshooting supplier and past crashing a pod to spot really nan supplier diagnoses nan problem.

“It’s rather nosy erstwhile nan point conscionable goes and looks astatine logs and metrics, and it’s like, ‘It looks for illustration this image doesn’t exist. There’s a typo. Go hole it.'”

As for his ain AI usage, Stefaniak is besides progressively utilizing these agentic tools, he said. 

“If you had asked maine six months ago, I would opportunity not much,” he explained. “In nan past six months, I honestly consciousness for illustration I’m moving much efficiently pinch immoderate of these tools. Some of nan really adjuvant usage cases I’ve been utilizing internally: BI. I utilized to person to spell fig retired SQL tables and SQL queries to spell fig retired what’s going connected pinch our customers and understand really they’re utilizing our work to make decisions. We person a BI supplier that tin now understand our tables and I tin inquire it questions that I’ve been wondering answers to for years, honestly, and it conscionable goes and figures it retired for maine based connected nan information that’s already there. That’s been a real, actual usage lawsuit internally that I’ve seen velocity up nan merchandise improvement process for us.”

Group Created pinch Sketch.