Post

You can run Mecatl on the desktop like any other harness. Or you can run the engine on Kubernetes where the agent loop is like any other application. All the ancillary systems (session storage, memories, tools, etc) can be run as separate services.

In addition to having optionality of implementation, these systems can be scaled, deployed and secured separately. We want to build a harness system (and larger ecosystem) that brings the power of modern agents to non-developers by shifting from the desktop to the cloud.

It is early days yet! It is rough around the edges. We are just getting started. But we have big plans. Read the blog post, kick the tires and let me know what you think.

This is super interesting. What size orgs are you targeting?

We want it to be useful in a lot of situations but commercially, Stacklok will look to build it into products aimed at large enterprises.

We’ve got about ten developers so we’re trying to figure out what it means to have isolated AI sandboxes and systems, either self hosted or not. This looks cool, but heavy infra with to get started.

The engine/UI split raises a useful failure test: disconnect the client just as a tool requests approval, then reconnect from another device. Does the pending request live in durable session state? That seems as important as replaying the conversation for the mobile clients you describe.

The tool calls all happen engine-side. But they can be distributed too so there are network failure modes to consider. I’m working on attaching to sessions during approval calls now. Having the client be able to re-sync there is the goal.

So yeah, the pending request state is in durable session state. And the goal is to have it be able to float based on leader election and allow for rolling upgrades and all while sessions are ongoing.