AI Runtime vs AI Gateway vs Agent Platform — kprompt, agentgateway, and kagent

Open a Kubernetes AI shortlist and three phrases collide: AI Runtime, AI Gateway, Agent Platform. Marketing uses “agent,” “MCP,” and “runtime” freely. Buyers type kagent, agentgateway, or “AI for Kubernetes” and land on demos that look similar in a screenshot. The jobs are not the same.

This hub separates the triangle used on this site: kprompt as an AI Runtime for cluster ops (PlanResult → approve), agentgateway as an AI/MCP/A2A gateway data plane, and kagent as a Kubernetes-native agent platform (Agents-as-CRDs). Hire by job. Deep dives live one click away.

One table — three layers

Same cluster can run all three. Confusing them buys the wrong control plane — or skips the only gate that prints a refuse-able plan before mutate.

AI Runtime — reason about the cluster, then ask before apply

An AI Runtime for Kubernetes (as we mean it) sits between human intent and the API server: gather context, reason, emit a typed plan with risk and blast radius, hard-deny wipe-class language, then apply only after approval. The laptop path is a CLI. The in-cluster path can Observe a namespace and page Slack — still without silent auto-heal.

$ kprompt "scale api to 3" -n staging

Plan
  1. scale Deployment/api replicas → 3

Risk: medium · blast radius: staging/api
Apply? [y/N]

• Best when: day-2 mutate, investigate/why/timeline, CI PlanResult gates
• Not when: you need Agents-as-CRDs or an LLM/MCP traffic proxy
• MCP role: kprompt mcp serve = IDE tool provider (read/plan-only), not a gateway

AI Gateway — front agent traffic, not the kubectl plan

An AI Gateway is a data plane for agentic protocols and LLM calls: route, authenticate, observe, budget, and audit hops between agents, tools, and model providers. agentgateway unifies ordinary HTTP/gRPC with LLM, MCP, and A2A on a Gateway API–shaped path. It does not compile “scale my Deployment” into a PlanResult.

• Best when: every LLM/MCP/A2A hop needs policy at the edge
• Not when: you only need a laptop NL ops CLI
• Confusion: “MCP” here means federating tool servers on the wire — not Cursor calling kprompt.plan

Agent Platform — agents as workloads next to apps

An Agent Platform treats agents as first-class Kubernetes resources: Agent CRDs, MCP tool catalogs, A2A composition, GitOps rollouts, mesh-aware governance. kagent (CNCF Sandbox, Solo.io origins) is that product. Platform teams own agent lifecycle the way they own Deployments. Human-in-the-loop is a feature of those agents — not the same artifact as kprompt’s PlanResult.

• Best when: “agents are a product we run on the cluster”
• Not when: you want a single gated day-2 CLI without agent control-plane ops
• Neighborhood: often evaluated next to agentgateway — platform vs gateway, still not ops-compiler

Why “runtime” collides

kprompt says AI Runtime. kagent says Kubernetes-native agent runtime. McKinsey ARK says Agentic Runtime. agentgateway is a gateway runtime for traffic. Same English word; four install shapes. We keep “AI Runtime for Kubernetes” for the ops/reasoning loop under plan → approve — and we link out honestly when the buyer’s job is gateway or Agents-as-CRDs.

Decision checklist

• Need a refuse-able plan before mutate / CI JSON gates? → AI Runtime (kprompt)
• Need LLM/MCP/A2A traffic policy on Gateway API? → AI Gateway (agentgateway)
• Need Agents as CRDs + MCP/A2A agent product? → Agent Platform (kagent)
• Need IDE assistants to call read/plan tools? → kprompt mcp serve (not a gateway)
• Need NL → kubectl chat fluency? → kubectl-ai (or kprompt for gated plans)
• Need analyzer-first “what is broken?”? → K8sGPT

Healthy coexistence

A realistic stack: agentgateway fronts LLM/MCP hops; kagent hosts internal agents behind that edge; operators use kprompt for gated day-2 and Observe for namespace paging. A kagent agent may propose a change that a human still applies via PlanResult. We do not ship that composition as one product — we claim the layers are complementary.

Agent Platform (kagent)     → author & run agents as CRDs
AI Gateway (agentgateway)  → govern LLM / MCP / A2A traffic
AI Runtime (kprompt)       → PlanResult → approve → apply (+ Observe)

Deep links by query

Honest limits

• kprompt is experimental — read every plan; prefer non-prod first
• kprompt is not an AI gateway, not Gateway API, not Agents-as-CRDs
• agentgateway is not a PlanResult ops CLI
• kagent is not a laptop plan-before-apply contract by default
• Shared MCP/A2A vocabulary does not collapse the three layers

If you leave this page with one sentence: gateways move agent traffic, platforms host agents, runtimes refuse bad plans before apply. Pick the layer that matches the ticket on your board — then open the deep dive for that corner of the triangle.