Factory Layer

What it does

Factory Layer turns a one-page product spec into a working TypeScript project (a CLI or an HTTP API) with one command and no human in the loop. A brain of Mozaik agents plans the work and derives acceptance checks from the spec, the spec's own examples first. Coding agents carry out each task through APB playbooks that build the code, run the package's quality gates and a reviewer side by side, and loop through fix rounds, escalating to a stronger model when needed. A Critic agent judges every finished task against the checks; rejected work is repaired or re-planned, accepted work is merged into main, and a clean copy is verified and delivered. Every run writes a report and a trace with the time, cost and models of each task, and Open Second Brain remembers past runs for the next plan.

How the agents run concurrently

At the start, the Architect (plans tasks) and the Acceptance analyst (writes acceptance checks) run as two concurrent Mozaik agent loops on the same spec. Their results meet in the shared run state, and the Critic then reviews the checks. Tasks whose dependencies are integrated are dispatched at once, so several APB playbook runs build different modules in parallel, each in its own git worktree. Inside every run, the quality gates (a script) and a reviewer agent run side by side and join into one verdict, and finished tasks are judged by Critic decisions running alongside the remaining builds. Coordination goes through one shared state with a single writer: brain agents only return typed decisions through submit tools, a scheduler starts runs as dependencies land, a lease bounds concurrent executor runs, an integration queue merges accepted tasks into main one at a time, and every transition is saved before its side effect, so a crash resumes cleanly. Recovery decisions (retry, repair, re-plan) are made per task while other tasks keep running, and a cutoff stops all live runs at once and delivers what is integrated.