I've been building a small open-source CLI called Lok for orchestrating multiple LLMs in a single workflow.

Instead of a single "agent", Lok treats models as interchangeable tools: one can plan, another can execute or critique, and others can verify output. The focus is on inspectable, composable workflows rather than autonomous behavior.

It's local-first, CLI-driven, and intentionally small. I've been dogfooding it for real development work and documenting what works and what breaks as I go.

The accompanying post is mostly about the mental model behind the tool: thinking of LLMs as infrastructure and control planes rather than chat interfaces: https://jakegoldsborough.com/blog/2026/introducing-lok-multi...

Code is here if you want to poke at it: https://github.com/ducks/lok https://crates.io/crates/lokomotiv

I'm curious whether this framing resonates with others experimenting with multi-model workflows.