Skip to content
NBS Workshop

Day 3 · Source + office library + memory

Wednesday 27 May 2026 · Tokyo · Engineers, in person

Day 3 is the handover. The substrate is open and source-available. The morning walks the engine and the library; the afternoon is build-your-own, with the option to wire Day 2’s generated hooks back into the SDK and test them.

Morning · ~3-hour block · Source, library, memory

Open the source. Walk the engine. Then go deep on memory — how agents organise their work and reach into enterprise data.

  • Orchestration runtime walk-through — engine files, wakeup decision tree, adapter engine
  • Office library walk-through — the post-it knowledge system, five perspectives per file, live provenance trail
  • Memory — two flavours:
    • Memory the agent builds about its own work — what it did, how it did it, what it learned. The agent reads its own previous work and prior chat history before acting.
    • Memory of the enterprise data the agent operates on — organised per use case, wherever the agent is deployed.
  • Context-window strategies — separate from memory. The LLM’s per-prompt token budget. History compression, when to delegate to subagents to keep the parent context clean, the ~20K-token tax per Claude subagent, working-directory hygiene.
  • Each engineer clones the repo and branches

→ Architecture & code · Office library · Memory

Afternoon · ~2 – 3-hour hands-on

Three tracks. Each pair picks one.

  • Track A — Wire and test the generated hooks. Take the hooks/*.ts files Agent 7 (Claude Hook) generated yesterday, import them dynamically into the SDK runner you built Day 1, register through options.hooks, and run test scenarios. Workshop 3 reference.
  • Track B — Build a tenth agent in code. A new agent for an in-house use case, wired into the existing nine-agent pipeline.
  • Track C — Build a custom front-end. Minimal HTML + fetch against the orchestration API. A different view of the same agents.

→ Workshop 3 — Test Generated Hooks · Setup · GitHub access

Why source-available matters

The front-end is one possible front-end. The team can build their own — a chat surface, a mobile app, a custom internal interface — pointing at the same orchestration. Business applications live in the front-end; the substrate stays the same.

How today feeds Friday’s MVP

The afternoon tracks are concrete MVP shapes. By end of day:

  • Track A engineers have a working hook-test runner — guardrail-shaped MVPs
  • Track B engineers have a new agent running locally — specialist-agent MVPs
  • Track C engineers have a custom UI talking to the substrate — front-end MVPs

Many engineers continue refining the same thread into Day 4 morning’s free prototyping time.

Materials

Next: Day 4 →