Why We Bet on Kubernetes for Agentic Workloads

There's a popular shortcut in agent infrastructure: glue an LLM call to a serverless function and call it a platform. It demos well. It falls apart the moment an agent needs to run for more than fifteen minutes, hold state across steps, or recover from a crash without losing its work.

Agents are infrastructure-shaped

A useful agent is not a stateless request/response. It is:

• Long-running — minutes to hours, sometimes days
• Stateful — it accumulates context, intermediate results, and memory
• Failure-prone — tools time out, models rate-limit, networks blip
• Concurrent — many agents, many tenants, shared resources

That list is a description of the workloads Kubernetes was built to schedule, isolate, and heal. We didn't pick it because it's trendy; we picked it because the alternative is rebuilding orchestration, retries, and isolation badly.

The "Lambda and a prayer" tax

// The seductive version — until step 12 of 40 the function times out.
export async function handler(event: AgentEvent) {
  const plan = await llm.plan(event.goal)
  for (const step of plan.steps) {
    await runTool(step) // no durable state, no resumability
  }
}

When that function dies at step 12, there is no resumable state, no retry boundary, and no trace of what the agent already did. You rebuild durability from scratch — which is to say, you rebuild a workflow engine.

What we use instead

On the K8gentic platform, the foundation layer leans on infrastructure that has already solved these problems:

A durable workflow makes the same agent loop look very different:

Every step is recorded, retried, and resumable. A crash at step 12 picks up at step 12 — not step 1, and not a support ticket.

The bet

Betting on Kubernetes for agents isn't nostalgia for YAML. It's the recognition that agentic is a systems-engineering problem wearing an AI hat — and the systems engineering was solved over a decade of running trillions of operations in production. We'd rather stand on that than on a prayer.