Skip to content
NBS Workshop

System architecture

The Living AI Office is the substrate underneath every agent system we ship. This section is the canonical architecture publication — the same documents partners and clients receive.

Start here — the platform in one diagram

The fastest way to see the whole system is the platform overview. Open it side-by-side with this section.

Platform overview Edge · Gateway · Compute · Runtime · Data. The full topology.
Hub-and-spoke view User Management as the routing hub. Six clients in, six destinations out.

The four choose-anything pillars

The enterprise-AI adoption thesis: you must be able to swap any of these four layers without rebuilding the others.

01 · Coding agent / SDK Claude Code · Pi Coding Agent · Claude Agent SDK · Codex · OpenAI Agents · LangGraph · custom
02 · LLM · any provider Anthropic · OpenAI · Gemini · OSS-80B. Any LLM on your own account, best-of-each per role
03 · Orchestration LIFEOSAI substrate or your own — same patterns
04 · Library + memory Three feeders · five perspectives · provenance backbone

Office library + memory — how the office remembers

The most under-named feature of agent systems is what they know on Tuesday morning before anyone speaks. Our answer: a five-stage pipeline of analyzers, story-builders, and final-story writers that wake every agent with full context.

Three feeders Office library · chat reflection · pinned facts → one memory folder
Multi-perspective Five lenses per file. The tea-receipt example.
Provenance Every claim traces to raw bytes

Worked examples

Reading architecture is easier with one real trace. Two are published in full.

One incident · five analyzers · one report An AI safety incident arrives Tuesday morning. Walk T+0 → T+next-day, all five stages.
What an agent SOUL looks like The Analyzer-CA spec. Identity, environment, lens, workflow.

Reference