Blog
Engineering notes from the team building Factory Nexus.
Latest
H2L in Omnimancer 0.3: a high model reads the repo and writes stories a blind worker can execute, a pool of small models does the edits with no exploration tools, and the high model scores each diff against its own acceptance criteria without ever reading the worker's explanation. Plus four things the first live plans got wrong.
Read post
What Omnimancer is (a headless coding agent for Claude, OpenRouter, DigitalOcean, Ollama and 13+ backends), the stream-json and exit-code contract an orchestrator builds against, how Factory Nexus runs it in swarms, and the five production failures that shaped it: narration exits, permission theater, timeouts, quadratic token cost, and rate limits that lost the whole session.
Read post
A tool-agnostic playbook for running several AI coding agents at once: the three isolation layers, when to use fan-out vs hub-and-spoke coordination, what parallelises and what doesn't, the five mistakes that cost us the most time, and an FAQ.
Read post
A complete git worktree setup for running AI coding agents in parallel — worktrees, port and database isolation, merging and cleanup — then the six failure modes we hit in production and the point where hand-rolling stops being worth it.
Read post
A precise definition of AI agent swarm coding, the four-layer architecture behind it (decomposition, isolation, verification, integration), an honest comparison of agency-swarm, Claude-Flow/Ruflo, Factory.ai and Factory Nexus, and a straight answer to Thoughtworks' caution rating.
Read post
How Factory Nexus runs parallel AI coding agents from the Agent Panel — orchestration and file lanes, worktree isolation, the three-reviewer gate, smart retries, and the integration verifier that runs before anything reaches a PR.
Read post
Why we're writing about parallel AI coding agents, git branch isolation, and AI code review — and what to expect from the Factory Nexus engineering blog.
Read post
