Hmm, a bit confusing, the models are not local. In this context, running on your desktop should meant he models, not the app, that is obvious.

This matches our experience. The biggest friction wasn’t agents being slow, it was agents doing too much between human checkpoints and breaking flow.

What worked for us was treating agents as step-scoped workers rather than continuous copilots. Each step had explicit inputs, allowed tools, and a clear exit condition, then everything reset.

Once agents couldn’t carry assumptions or permissions forward, they became much easier to run alongside normal dev work without surprises. Curious how you’re handling step boundaries and “when to stop” signals.