+ Skills, tools and agents are Kubernetes resources. Write them, apply them, review them in a pull request. No framework to + learn, no glue code to maintain. +
+
- $ {INSTALL_TABS[tab].cmd}
-
- - kagent runs your agents where your workloads already live — on Kubernetes. - Deploy, observe, and govern AI agents with the tools your platform team - already trusts. Open source. Production grade. Built by the founders of Istio. -
-You already know how to run things on Kubernetes — pods, services, RBAC, observability, GitOps. kagent makes agents another first-class workload, not a new platform to learn.
-{body}
-Every feature works with a single helm install. No add-ons, no extra databases, no waiting for enterprise.
Multi-LLM, multi-framework, multi-tool. No lock-in at any layer of the stack.
-From day-zero installs to platform-wide rollouts, kagent powers the agent surface area inside platform engineering organizations.
-{body}
-kagent doesn't reinvent the wheel. It composes the protocols that the agent ecosystem is converging on — so you can swap models, frameworks, and tools without rewriting your agents.
-Connect any tool — REST, gRPC, databases, internal APIs — through a single declarative spec.
-Compose multi-agent systems with first-class delegation, hand-off, and shared memory.
-Every prompt, tool call, and token emits OTel traces — the same pipeline as the rest of your stack.
-Agents, sessions, and tools as CRDs. GitOps, RBAC, and admission controllers come for free.
-+ Thousands of agents per cluster, resuming in milliseconds, sandboxed by default. kagent runs on Agent Substrate — the + runtime built for what agents become next. +
+kagent is a CNCF Sandbox project. Roadmap is public. PRs are welcome. The maintainers hang out in Discord and answer in hours, not days.
-Apache 2.0. Public roadmap, public RFCs, public weekly community calls.
- github.com/kagent-dev → -Maintainers and operators sharing patterns, debugging clusters, shipping integrations together.
- Join Discord → -From individual operators to teams at hyperscalers. Weekly releases, monthly community demos.
- See contributors → -+ Agents resume in milliseconds, hold nothing while idle, and run sandboxed from the first instruction. Substrate's core is + built on Kubernetes, so this is the platform you already operate. +
++ Agents don't usually need to run 24/7, so reserving constant compute for them is very inefficient. Suspend an agent and + it holds nothing. Resume it and it is back before anyone notices. +
+One Helm chart. One CRD. One namespace. You'll be looking at agent traces before your coffee gets cold.
-+ Agents share pools of workers instead of getting their own dedicated compute. Substrate packs more agents onto the system + than it has compute for, because agents that aren't running don't hold any. +
++ Agents can be very dangerous. All compute inside Substrate runs in a sandbox runtime, and Substrate controls all network + traffic in and out of the agent — nothing gets in or out without explicit permission. +
+