↓Skip to main content

About

Built by operators, for operators.

The Human #

Josh VanDeraa - Services Director at Network to Code, author of Open Source Network Management and Network Automation with Nautobot. 15+ years in network engineering and automation. Building Nightwork Labs to prove that AI agents can ship real products, not just answer questions.

The Agents #

Nightwork Labs runs on a team of AI agents coordinated by Harry (CEO/PA). They work autonomously - overnight builds, hourly task loops, nightly security scans, daily product reviews.

  • Harry 🎩 - CEO. Strategy, delegation, final decisions.
  • Hermione 📚 - COO. Code review, quality, operations. Doesn’t let anything slide.
  • Ron 🔄 - VP Operations. Task orchestration, pipeline management.
  • Dobby 🧦 - Lead Engineer. Writes the code. One task at a time, every time.
  • Scout 🔍 - Research. Web research with no write access, by design.

This isn’t a gimmick. The agents built this website, review every product update, and run the morning briefing. Harry drafted the tweet thread promoting our products. Dobby built the Grafana dashboards in the observability kit. Hermione catches bugs Dobby misses.

How We Work #

Every product starts the same way:

  1. Harry identifies the need from running our own agent stack
  2. Hermione writes the PRD and reviews it with synthetic customers (Morgan, Alex, Dana)
  3. Ron breaks it into tasks and manages Dobby through the build
  4. Hermione reviews the output - max 5 rounds until approved
  5. We ship it, then our reviewer personas tear it apart so we can improve it

The pipeline runs 24/7. Products get daily updates for the first two weeks after launch.

What We’ve Learned Running Agents in Production #

  • Q4 quantization on 122B models causes hallucinations in code review. Q8 fixed it immediately. That insight is in the Local LLM Setup guide.
  • 60% of installed agent skills go unused after 30 days. The staleness dashboard in the Observability Kit catches this.
  • Discord without structure becomes noise. The right channel layout turns it into an ops dashboard. That’s the Discord Setup guide.
  • Alerts without destinations are decoration. Every alert needs to fire somewhere real (Discord webhook, not just Grafana).

These aren’t theoretical observations. They’re production scars, and every one became a product feature.

The Stack #

Running on a Mac Studio (M2 Ultra, 225GB unified memory):

  • Qwen3.5 122B-A10B Q8 for review and reasoning
  • Qwen3 Coder 30B for implementation
  • Qwen3.5 27B for ops and routing
  • Prometheus + Grafana for monitoring
  • OpenClaw for agent orchestration
  • Discord as the ops dashboard

Total local inference cost after hardware: $0/month for the majority of agent workloads.

Nightwork Labs is a subsidiary of CMJJV.