I compared AI software factories with agent control planes while researching how coding agents can work through a GitHub issue backlog. I wanted to understand what changes when that work is split across worktrees and remote machines.
Encore's architecture write-up helped separate the delivery loop from the rules governing it. The loop takes a task into an isolated environment, builds and validates a change, then moves it towards deployment. The control plane holds configuration, permissions and deployment policy. The worker operates within those rules.
This is a conceptual sketch of the separation: solid arrows show the work moving forward; dotted arrows show the policy constraining it.
Detent makes the state transitions concrete. Its documented workflow uses a GitHub Projects board to move an issue through a worktree, validation and human review, with a serialised merge lane. Sandcastle's workflow keeps its durable state in git and GitHub and reconstructs it on the next run. It leaves the target-branch merge to a person. Those are useful design choices to compare, not guarantees I've tested.
My takeaway is to ask where task ownership and the next allowed transition live before adding workers. A worktree gives a worker somewhere to edit; it doesn't by itself decide who owns an issue. Where that shared ownership should live across machines remains open for me. This was research, with no installation or recovery test.