Build1 publisherNot yet confirmed elsewhere3 min readPublished
Kubernetes creators' Mecatl harness splits the coding-agent loop from its tools and session state
Kubernetes creators Craig McLuckie and Joe Beda's startup Stacklok is building Mecatl, a coding-agent harness designed to run in the cloud. It pulls tool calls, session state and memory out of the agent loop so each can be put into systems an operator can manage.
The Engineer · Build desk
What happened
- OpenAI and Anthropic have been trying since 2025 to move their own agent harnesses into the cloud, with varying levels of success according to Latent Space.
- Beda says desktop-first harnesses often start the agent loop, local execution and session state inside the same process.
- Mecatl began in June as an open source project on GitHub, and Latent Space calls it the most interesting of Stacklok's products.
- Stacklok raised a $17.5 million Series A in 2023 from Accel, Madrona and Bain Capital, then moved from software supply chain security to Kubernetes-based agent products.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- capability Keeping session state and memory outside the loop is what would let a Kubernetes-style controller stop, move or pause an agent without losing its context.
- decision Teams already running desktop harnesses inside VMs, containers or sandboxes have a specific test to apply: whether an agent can be paused safely while it waits for a human.
- exposure Code and context now held on developer laptops would move into whatever state store and execution environment a team plugs in, putting what Beda calls the IP under central control.
A coding agent is an LLM calling tools in a loop while it carries the needed context forward [1]. The easiest place to build that is one machine. Many agents began as terminal tools and grew into desktop apps, and Claude Code started as a CLI [2]. In Beda's account, plugins do not change that shape. "So even if they have pluggability, the architecture is still one process, or one set of tightly coupled processes, that are built to run on the desktop," he said [15].
Mecatl keeps the agent loop independent of the client, the model provider, the state store and the execution environment [17]. Beda called the loop "an application that we know how to run in the cloud well" [18]. The sensitive parts come out of it. "What if we take that and we separate that out from the more sensitive operations, like tool calling and bash, and then also separate out things like session management and memory, so that these things are no longer stored as JSONL files sitting on disk, but can natively be put into manageable systems," he said [18].
Kubernetes uses control loops to keep applications running, recover them from failures and scale them [5]. I think one condition has to hold before that pattern transfers to agents. A controller has to be able to stop the loop and start it somewhere else without losing the session. A session kept as JSONL files on a local disk [18] is tied to that disk. Putting session state and memory in a separate store is the precondition for any reconciliation. It is also the part of Mecatl that McLuckie and Beda describe most concretely [17].
Pausing is where I find the argument strongest. Beda said moving a desktop harness into a VM, container or sandbox tends to cause trouble with the agent's lifecycle, and with quiescing it: pausing it safely while it waits for human input [16]. I'd expect a waiting agent in the decoupled design to be a record in a store, with no process held open for it. "Our thinking is, well, what if we rethought the harness from the ground up to actually run in the cloud?" Beda said [20].
Latent Space's list of what makes the cloud move hard starts with keeping sessions reliable, then isolating tool execution and preserving context [4]. McLuckie and Beda made their case in management terms. McLuckie described the aim as "moving the locus of value from the desktop to the cloud" [11]. "That's where the IP sits," Beda said. "And if you want to be able to actually bring that under management, it's that much more difficult when it's sitting on a desktop." [12] McLuckie also asked: "Why is the agent loop so intricately coupled to the tool calling subsystems? Why is agent identity being reasoned about through the lens of human identity systems?" [13]
The name is pronounced MEH-kah-tl, an Aztec word for a cord or rope. Latent Space's writer first guessed "me-cattle", a pun on the Kubernetes rule of treating servers like cattle, not pets [10]. The interview does not describe how Mecatl recovers a failed session. Nor does it compare Mecatl's reliability with desktop harnesses such as Pi, a "minimal agent harness" that projects like Flue build on [19].
What to watch
- Whether Mecatl documents how a failed or paused session is restored from its state store, and which stores it supports.
- Whether the cloud harnesses OpenAI and Anthropic are building adopt a similar split between loop, tool execution and session state.
- Any published reliability comparison between Mecatl and desktop-first harnesses such as Pi.