Canon

04 — AI Workbench, Twin, Spine Agents and AI Continuity Memory

Canonical reference: 00-SPINEWORKSPACE-INTEGRATEWISE-CANON.md §8, §9, §4.3
Status: Canonical

1. AI Workbench definition

The AI Workbench is the machine-side operating environment in which Twin and Spine Agents perform business work over the shared Business Context Spine.

It does not need to appear to the customer as a separate visible app. It is first-class architecturally because an AI worker needs a stable operational environment just as a human does.

The AI Workbench exists to reason over and act on Business Context. Without Business Context, the AI Workbench has nothing to work with. With it, the AI Workbench already knows the business objective, the current state, the evidence, the authority, and what needs to happen next.

2. What the AI Workbench contains

  • active objective
  • assigned Work record and current state
  • authorized Business Context Bundle
  • relevant Business Memory
  • AI Continuity Memory
  • relevant entities and relationships
  • evidence and provenance
  • current plan and subtasks
  • dependencies and blockers
  • agent roster and agent state
  • available capabilities and provider health
  • authority/delegation envelope
  • proposals and pending approvals
  • execution state
  • expected versus observed outcomes
  • verification, exception and handoff state

3. Twin versus AI Workbench

Twin is not the entire AI Workbench. Twin is the contextual co-worker closest to the human: it retrieves, reasons, explains, prepares and proposes.

The AI Workbench can host or coordinate multiple workers behind Twin, including specialist agents, research agents, execution agents, verification agents, workflows and model calls. Twin may perform work itself or assign sub-work through governed agent/capability paths.

4. AI Active Context

AI Active Context
= Business Context Projection
+ AI Continuity Memory
+ Current Objective
+ Current Work State
+ Evidence
+ Authority Envelope
+ Available Capabilities
+ Agent Runtime State

The invariant:

Shared business reality; separate participant continuity; governed promotion between scopes.

5. AI Continuity Memory

AI Continuity Memory is the AI-side continuity layer. It exists so Twin and Spine Agents do not restart from zero on every turn, model change, session or handoff.

It can hold: prior assignment state, working summaries, unresolved questions, current plans, active hypotheses, tool/capability state, agent handoff context, prior observations, pending verification, interaction preferences and model-independent continuity information.

AI Continuity Memory is not organizational truth. An AI-generated inference, summary or hypothesis becomes organizational memory only after the appropriate Triage, Governance and governed-write path.

Constitutional rule

AI Continuity Memory ≠ Organizational Truth

6. AI continuity must survive model changes

The AI Workbench, Work state and AI Continuity Memory are durable platform concepts. A specific model is a replaceable reasoning provider.

A model resolver should select by task profile, context requirement, reasoning effort, modality, latency, cost, policy and availability. The model contract should use stable profiles such as fast, reasoning, code or specialized, rather than hard-coding one vendor/model version into Twin.

7. AI context pipeline

Business Context Spine
+ AI Continuity Memory
+ Objective
+ Current Work State
+ Evidence
+ Authority
+ Capabilities
+ Runtime Health
        ↓
AI Context Assembly
        ↓
AI Workbench Context
        ↓
Twin / Spine Agent Runtime

Human and AI share the same underlying business reality without requiring identical representations.