Skip to content
NBS Workshop

Day 2 · Orchestration + the nine agents

Tuesday 26 May 2026 · Tokyo · Engineers, in person

The pivot day. Day 1 taught one agent in a terminal. Today shows nine working together as an office — and engineers build the pipeline together by orchestrating Claude Code, then verify that the guardrails the agents generate actually work.

Morning · ~2-hour block · Build together

The nine-agent pipeline is built together, not pre-built and shown. The public repo github.com/metaweavehq/ai-guardrail-lab ships the ten agents as instruction files only — no source code in the agent folders. Engineers paste prompts into Claude Code; Claude reads the instruction files and orchestrates the pipeline live.

  • Why orchestration — failure modes when scaling from two agents to twenty
  • Architecture answers — orchestration substrate + office library
  • Engineers clone the repo and run the pipeline themselves, paste-by-paste
  • Highlight: agents #6 and #7 produce the actual guardrail code (Policy-as-Code and Claude Hook)

Alongside the build, the same system is shown running in production through the hosted platform UI — so engineers see both the build experience (terminal, instruction files, prompts) and the deployed experience (full orchestration, multi-company, audit trail) side by side.

→ Workshop 2 — Run with Claude Code Terminal · The nine agents

Afternoon · ~3-hour hands-on

Browser logins. Modify agents. Test the guardrails.

  • Twelve web logins distributed at session start
  • Each pair opens an agent in the UI, edits Skills and Instructions, re-runs
  • Run the generated guardrails against the source incidents — confirm prevention behaviour
  • Pair-design a new guardrail if time permits

→ Setup · web platform access

How today feeds Friday’s MVP

By end of day, each engineer has:

  • Run the nine-agent pipeline themselves (instruction-driven, paste-by-paste)
  • Modified at least one agent in the platform UI
  • Tested that a generated guardrail actually catches what it was designed to catch

Several of these become MVP candidates for Friday’s presentation — a modified agent for a specific in-house use case, a new guardrail designed and tested for a particular class of risk.

Materials

Next: Day 3 →