EnoughFactory

A public prototype, no longer in development

I built EnoughFactory to explore coordinating coding agents across my own machines. It remains a public prototype, and I’m no longer working on it. The source is available to study or adapt; the sections below describe the work explored in the prototype.

Define the finish, then divide the work

A goal gives EnoughFactory a desired outcome in a repository. The plan turns that outcome into tasks, prerequisites, and checks. Before independent work begins, shared interfaces and dependencies can be recorded so one task knows what it needs from another.

The board and dependency graph make that plan visible while it runs. Independent tasks can proceed together; dependent tasks wait for their inputs. Scheduling considers available devices, worker limits, estimated resources, and tasks expected to change the same files. Those estimates guide placement, rather than acting as hard resource quotas or restrictions on which files an agent can edit.

Parallel attempts, one combined result

Each task gets an isolated workspace, using Git or ArtifactFS according to the environment. Its conversation, attempts, terminal, preview, and checks stay attached to the work. Agents can make progress independently without sharing a single checkout for their edits.

Bringing the candidates together is its own stage. Changes are integrated serially and checked against the combined result. If a task needs repair, its replacement and affected dependants can be rebuilt while unrelated work is retained. Execution whose status is uncertain is reconciled before replacement work is dispatched.

An answer is the start of the review

An agent finishing a turn and a task being accepted are separate events. Task-level checks assess an individual output; plan-level checks assess the integrated goal. Captured commits, prerequisites, attempts, and integration evidence provide a record of how a result reached that point.

The workspace keeps that evidence next to the conversation and code. You can inspect changes, open a terminal, view a preview, and steer the work as the goal develops. Manual, assisted, and autonomous operation describe how much coordination the app performs; agent approval settings remain a separate decision.

Work runs on the machines you own

EnoughFactory connects Mac and Linux devices, with conversations staying on their owning device. A device service continues independently of the desktop window. The app manages its own container runtime: a private Lima virtual machine on Mac and a rootless Docker engine on Linux.

The public prototype has MIT-licensed source. It required users to bring their own accounts and model usage for the supported coding agents. Device pairing uses authenticated connections with an encrypted fallback, and signalling can be self-hosted. The result is a workspace for coordinating your machines as well as your tasks.