Latest

AI Agents

Mecatl, AI agents arrive at Kubernetes

The creators of Kubernetes release Mecatl, an open source agent harness that separates the agent loop from the client, the model and the environment.

Mecatl, AI agents arrive at Kubernetes
Mecatl, AI agents arrive at Kubernetes

Most coding assistants live where the developer lives. The loop that calls the tools, the conversation state and command execution all run on the same laptop, and that poses a problem that does not appear until the organization tries to deploy the tool for thousands of engineers instead of six.

That is the diagnosis put forward by Craig McLuckie and Joe Beda, two of the creators of Kubernetes, from their company Stacklok. Their answer is Mecatl, an open source agent harness that launched in June 2026 and is designed to run on Kubernetes instead of on a desktop. McLuckie laid out the full argument on the CNCF blog on September 28 and the two later expanded on it in an interview with Latent Space published on October 7.

The thesis rests on a historical comparison its authors know well. The problems of maintaining reliable sessions, isolating tool execution and preserving context look quite a lot like container orchestration in the early part of the last decade, the ground from which Kubernetes emerged in 2014. If agents have the same kind of problems distributed applications had, the bet is that they will be managed with the same tools.

Why doesn't a desktop harness scale?

The starting point of the argument is architectural, not about performance. In a desktop harness, the agent loop, local execution and session state start inside the same process, as Beda describes it. Even if the tool offers plugins and some modularity, the architecture is still a single process or a set of tightly coupled processes designed to run on a personal machine.

That decision has consequences when enterprise deployment arrives. An agent's code and context are often derived from the local environment, and that is where the intellectual property of the work sits. Beda summed it up in one direct sentence: governing something that lives on a desktop is far harder than governing something that lives in the company's infrastructure.

It is worth asking whether packaging the harness into a virtual machine or moving it into a container is enough. The two founders argue that this direct transfer, which they call *lift and shift*, drags lifecycle problems along with it.

The specific case they mention is the safe pause: stopping an agent while it waits for a response from a person without losing the thread of what it was doing. An architecture that never anticipated that intermediate state does not solve it by adding a virtualization layer.

Mecatl a single proposal

Mecatl's proposal is to break the harness into pieces that can be managed separately. The agent loop is kept independent of the client, the model provider, the state store and the runtime environment. In Beda's words, it is about taking the part of the system that is known to run well in the cloud and separating it from the more sensitive operations, such as calling tools or accessing a shell.

The most tangible change is where the session lives. In desktop harnesses, the state of a conversation with an agent often ends up as a collection of JSONL files on disk. Mecatl moves that state to managed systems, and that opens the door to querying it, auditing it, moving it between machines and applying policies to it without depending on someone copying a folder.

Piece In a desktop harness In Mecatl's design
Agent loop In the same process as the client Independent application in the cluster
Session state Files on the local disk Managed and queryable store
Tool execution On the user's machine Isolated from the agent loop
Model provider Fixed in the local configuration Swappable without touching the loop
Pause waiting for a person Not anticipated in the architecture First-class intermediate state

McLuckie adds to that design a deeper question about identity. His criticism is that an agent's identity is being reasoned about with identity tools designed for people, and that inheritance shapes the rest of the design. It is an observation that points to a problem regulatory organizations will hit before others, because they need to know who acted and on whose behalf.

ToolHive and the gateway to agents

Mecatl is not the first piece that Stacklok has released as open source. Before it came ToolHive, a platform for running and governing MCP servers, the protocol agents use to connect to external tools.

It began as a way to manage those servers in Docker containers on the desktop and has evolved into a gateway on Kubernetes, with a registry and an operator that helps deploy them and handle inbound and outbound authentication.

The sequence makes sense. First you solve how the agent reaches the tools, then how the agent itself runs. The third piece is a model gateway, the AI Gateway, which is the only thing the company has not yet open sourced and which Beda places on the roadmap with the intention of releasing it.

The functions that gateway covers define well what its customers care about. Access control, budgets and routing between model providers, plus periodic consumption reports per team. What it does not do, by design, is choose the model based on the task. McLuckie explained that they prefer that semantic choice to happen inside the harness, where more context is available to decide. It is a notable difference from other products on the market that sell precisely intelligent routing as their main feature.

The business model behind the open source

The obvious question facing two open projects is where the money comes from. McLuckie's answer is that ToolHive and Mecatl are meant to be useful on their own, and that the company sells what it calls the enterprise backbone: a control plane that provides coherent identity, authorization, policies and auditing over the open pieces.

That split is not new in the infrastructure world, and it explains why the founders' pitch insists so much on the parallel with Kubernetes. The comparison they make is explicit: just as Kubernetes allowed companies to run in the cloud without tying themselves to the particularities of a specific provider, their stated goal is to reduce organizations' dependence on frontier labs and hyperscalers. The tools singled out are Codex, Claude Code and GitHub Copilot, among others from the big providers.

The motivation has a clear commercial component. If an agent's value stays on the employee's laptop, the company cannot govern it or accumulate it. If it moves to its infrastructure, it can.

Beda illustrated it with an analogy any IT manager understands: an employee's conversations with an agent, while working for the company, belong to the company, and they will want to be managed and governed the way corporate email is managed.

The bet that agents will be managed like email

The idea underlying Mecatl is less technical than organizational. Over the past decade, companies learned to treat their infrastructure as a centralized, governable resource, and that learning is what turned Kubernetes into a standard. Applying the same reasoning to agents means accepting that an agent's work is a company asset and not a private conversation on a machine.

That shift has a cost the founders acknowledge in passing and that deserves a direct look. A centralized harness is more governable and also more observable by the organization that deploys it. For the developer using the agent, the convenience of having everything on their laptop does not disappear by decree, and many tools will keep living there because they solve the immediate problem well.

What Stacklok is testing is whether there is a market willing to pay for the governable version before an incident makes it mandatory. With two open projects and a control plane as its product, the answer will come from the sectors that already treat their data with that level of formality.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading