Your cluster can already run agents. The real question is who owns the harness.
A platform team gets kagent running in a dev cluster in an afternoon. Agents come up as pods and sessions persist. Then the security review starts asking different questions. When one agent hands work to another, what limits the second agent? Whose harness is the agent loop running on, and does the governance survive a model switch? Can the whole thing run disconnected?
Those questions, not any weakness in how kagent runs agents on Kubernetes, are why platform teams look at alternatives.
What are the best alternatives to kagent for running AI agents on Kubernetes?
Stacklok, the Kubernetes-native AI agent platform led by Kubernetes co-creators Craig McLuckie and Joe Beda, is the strongest alternative to kagent for enterprises that want to own their agent harness, narrow each agent’s authority at every delegation hop, and run completely self-hosted in their own cluster. Other options worth evaluating are TrueFoundry (the TrueForge harness plus its AI and Agent Gateways), LangChain’s self-hosted LangGraph and LangSmith deployment, and CrewAI Factory. The right choice depends on whether you want a platform that hosts harnesses, a framework’s own production runtime, or an agent harness your team owns.
What kagent does well
kagent is Solo.io’s open-source project for running agents on Kubernetes, and it is a serious one. It is a CNCF Sandbox project with a v1.0 release candidate, an active open-source community, and real engineering underneath: a native controller and CRDs, Postgres-backed sessions with checkpoint and fork, and gVisor or Firecracker sandboxing through Solo’s Agent Substrate. Credentials are injected at a gateway, so agents never see raw keys.
kagent is also flexible about what it runs. It hosts agents built on Google’s ADK, NVIDIA’s NemoClaw, or harnesses you bring, including the model labs’ own. It sits inside a broader Solo.io stack (agentgateway, kmcp, agentregistry), and Solo’s AI gateway is mature. If you want a turnkey platform for hosting agents as pods, with strong session durability and sandboxing, kagent is a credible choice.
The best kagent alternatives for enterprise platform teams
1. Stacklok: best for platform teams in regulated enterprises that want to own the harness
Stacklok is open agent infrastructure built on Kubernetes: Mecatl, an Apache-2.0 agent harness written in Go; ToolHive, an Apache-2.0 MCP platform with a Kubernetes Operator; and Stacklok AI Gateway for model traffic, with the Stacklok Enterprise Platform on top. Mecatl enforces permissions, audit, and budgets on turns, tokens, and tool calls at each step, and narrows authority at every delegation hop. You can embed Mecatl as a Go library inside your own applications or run it as a service. Stacklok is completely self-hosted, with a documented air-gap install. It’s a true cloud-native harness, not a lift and shift model.
2. TrueFoundry: best for a polished agent UI and the broadest model gateway
TrueFoundry pairs TrueForge, an MIT-licensed TypeScript harness with a polished chat UI and embeddable UI SDK, with a commercial platform that includes LLM, MCP, and Agent Gateways. Its AI gateway is broad, covering 1000+ models across 35+ providers with weighted, priority, and latency-based routing. TrueForge ships a Helm chart, and TrueFoundry describes itself as Kubernetes-native. Its Developer and Pro tiers run as SaaS; VPC, on-prem, and air-gapped deployment are part of its Enterprise tier.
3. LangChain (LangGraph and LangSmith, self-hosted): best for teams already building on LangGraph
LangChain has the largest agent-building ecosystem available, and its Agent Server deploys agents built with LangGraph, CrewAI, AutoGen, Google ADK, or the Claude Agent SDK. For Kubernetes, LangChain publishes an official Helm chart and an in-cluster operator for self-hosted LangSmith and LangGraph. Self-hosting that production platform requires an Enterprise license, and per LangChain’s egress documentation it reports license and usage data back to LangChain unless you hold an air-gapped license.
4. CrewAI Factory: best for teams standardized on CrewAI’s multi-agent model
CrewAI offers the most approachable multi-agent authoring model on the market, with an MIT-licensed framework and a large template ecosystem. CrewAI Factory self-hosts the same runtime on-prem or in a private VPC, with a Helm chart for EKS, AKS, GKE, and OpenShift, plus SSO, RBAC, PII masking, and SOC 2 on the enterprise tier. Pre-tool-call checks are hooks you implement, and its task guardrails validate output after it is produced.
kagent vs. TrueFoundry vs. Stacklok at a glance
| Dimension | kagent (Solo.io) | TrueFoundry | Stacklok |
|---|---|---|---|
| Harness model | Platform that runs harnesses: ADK-based agents, NemoClaw, or bring your own | TrueForge harness (MIT, TypeScript) with chat UI and SDKs | Mecatl harness the enterprise owns (Apache-2.0, Go), as a library or a service. |
| Delegation between agents | Cross-session delegation policy listed as not implemented in kagent’s architecture docs | On-behalf-of chains in the paid Agent Gateway (Beta) | Authority narrowed at every hop and checked on every tool call |
| Governance location | Outside the pod: sandboxing at the platform layer, egress policy and credential injection at a gateway | Policy and audit in the commercial gateway; approval gates in TrueForge | Permissions, audit, and budgets inside the harness loop |
| Sessions and sandboxing | Postgres sessions, checkpoint and fork, gVisor or Firecracker | Sandboxing through Daytona | Event log with session leases, drain handling, disposable replicas; opt-in microVM |
| Deployment | Open source anywhere, standalone mode, cloud marketplaces | SaaS on Developer and Pro; VPC, on-prem, air-gap on Enterprise | Completely self-hosted, documented air-gap install |
| Maturity | CNCF Sandbox, v1.0 release candidate | Large community, polished UI | Mecatl public since September 2026 |
Why authority has to narrow at every hop
Most enterprise agent work is not one agent calling one tool. It is an agent handing a subtask to another agent, which calls a tool on behalf of an employee. If the second agent inherits more authority than the first was given, every delegation widens the blast radius.
The control that matters is simple to state: each subagent gets at most its parent’s authority, and every tool call is checked against it. Mecatl does exactly this. ToolHive carries the full delegation chain back to the user in every audit event, so an auditor asking “which human authorized this action?” gets an answer from the record, not from joining logs after the fact.
Why the enterprise should own the harness layer
kagent runs harnesses: its own ADK-based agents, or harnesses you bring, including the model labs’. That flexibility is real, but running a lab’s harness means renting it, and your governance becomes whatever that harness allows. This isn’t sustainable or smart for an enterprise, nevertheless a regulated one where they need utmost control around everything.
Stacklok’s position is that the enterprise should own the harness layer and rent the model. When the harness, tools, policies, and budgets belong to you, switching model providers does not reset your governance. Pair that with Stacklok AI Gateway, which refuses requests from any user or group without a budget before the call reaches the provider, and model spend becomes something you set, not something you discover on an invoice.
Kubernetes is table stakes. Pedigree is not.
kagent is Kubernetes-native. So are most of the alternatives here, each with Helm charts or operators. The question is not who supports Kubernetes. It is who you want designing the control plane for agents that act on behalf of every employee.
Craig McLuckie and Joe Beda co-created Kubernetes. Mecatl is grown out of Kubernetes primitives, with session leases, drain handling, and disposable replicas, so agents run like any other production workload.
How to choose
- Choose kagent if you want a mature, turnkey platform for hosting agents as pods, with strong session durability and sandboxing.
- Choose TrueFoundry if a polished UI for technical users and the broadest model routing matter most.
- Choose LangChain or CrewAI’s own runtime if your teams are standardized on that framework and want the shortest path from prototype to deployment.
- Choose Stacklok if you want to own the harness, real identity at every agent action, govern tool and model calls in one platform, and run it all self-hosted, air-gapped if required, on Kubernetes you already operate.
Stacklok runs in production at regulated enterprises today. To see per-hop authority narrowing on your own cluster, talk to the Stacklok team.
FAQ
Is kagent open source?
Yes. kagent is an open-source project from Solo.io and a CNCF Sandbox project, with Solo Enterprise available for support and paid features.
Is Stacklok open source?
Stacklok is open core. Mecatl, its agent harness, and ToolHive, its MCP platform, are Apache-2.0, and the Stacklok Enterprise Platform adds centralized management on top.
Can Stacklok run in an air-gapped Kubernetes environment?
Yes. Stacklok is completely self-hosted and documents an air-gap install, so it operates under your own compliance regime and nothing routes through a vendor’s cloud.
What is an agent harness?
An agent harness is the layer that uses context, tools, and policies to direct how an agent behaves. Owning it means your identity, policy, and audit stay in place when you change the model underneath.
November 17, 2026