What it is: A livestream where Cole builds a "dark factory" — a system where an AI coding agent writes code, opens PRs, reviews them itself, merges, and loops back for the next task, with zero human code review.

How it works:

• Built on top of Archon, Cole's open-source harness builder for AI coding
• Archon workflows are node-based: each node is either a prompt sent to a coding agent session or a deterministic command (validation, testing, context curation)
• The dark factory chains these nodes into an autonomous loop: plan → implement → test → review → merge → repeat
• Human approval gates are optional — Cole is deliberately removing them to test full autonomy
• Archon handles orchestration, parallel agent sessions, and logging so you can monitor what's happening

Tools & links:

• Archon — open-source harness builder for AI coding agents, the core engine behind the dark factory
• Archon workflows — prepackaged workflows for fixing GitHub issues, creating PRs from ideas, PR validation/review, and PRD generation with human-in-the-loop
• Archon ships as a Claude Code plugin with a built-in skill — say "use archon to build this feature" and it picks the right workflow

How mx-workflow could steal from this:

• Archon's node-based workflow model (prompt nodes + deterministic command nodes) is a more structured version of what mx-workflow does with slash commands — consider whether mx-workflow commands could be composed into repeatable, chainable workflows
• The dark factory pattern (write → test → review → merge → loop) is essentially `/mx:shipit` on autopilot — Archon adds the orchestration layer that makes that loop autonomous
• Archon's approach of enforcing deterministic validation steps (not leaving linting/testing to the agent's discretion) directly mirrors mx-workflow's `/mx:validate` philosophy but makes it structural rather than optional
• Worth studying how Archon separates "things the agent decides" from "things the harness enforces" — that distinction is the whole game for reliable AI coding

What it is: GLM-5.1 from Z.ai (Zhipu AI), an open-weight model released under the MIT license that's outperforming GPT-5.4 and Opus 4.6 on software engineering benchmarks.

Benchmarks that matter:

• SWEBench Pro (real-world software engineering): GLM-5.1 scores 58.4 vs GPT-5.4's 57.7 and Opus 4.6's 57.3
• Beating closed-source frontier models while being fully downloadable is the headline

Why this matters for web app builders:

• MIT license = download, fine-tune, run locally, embed in your pipeline, no usage fees
• Available on HuggingFace right now — model weights are public
• You could fine-tune this on your own codebase patterns (SvelteKit, Supabase, your conventions) and get a coding model that knows your stack intimately
• Potential to plug into local AI coding setups alongside tools like Archon or Claude Code as an alternative model backend

Tools & links:

• GLM-5.1 on HuggingFace — THUDM is Zhipu's HuggingFace org, search for GLM-5.1 weights
• Z.ai (Zhipu AI) — the Chinese AI lab behind the model

Bottom line: The gap between open-weight and closed-source models on coding tasks is effectively zero now. If you're building web apps and paying per token, this changes the math.