07 · Twin Workbench
The agentic workspace — reasoning, planning, work allocation, agent delegation, and execution for one user
The agentic workspace.
The Twin Workbench answers one question:
What can I do for this user, what should the user do, and what should be delegated?
It is the AI counterpart to the Human Workbench. One Twin per user, scoped to their role domain. The Twin reasons, plans, allocates work, delegates to agents, executes under governance, and reconciles outcomes.
Twin Anatomy
USER
│
└── ONE TWIN
│
├── Understand context
├── Observe work
├── Analyze
├── Plan
├── Allocate work
├── Delegate
├── Execute
├── Communicate
├── Verify
├── Reconcile
└── Report
Core Capabilities
| Capability | Description |
|---|---|
| Understand context | Assemble context from the Spine via Projections |
| Observe work | Monitor the Work Graph for relevant changes |
| Analyze | Reason about situations and options |
| Plan | Propose step-by-step execution plans |
| Allocate work | Assign work to humans, agents, or self |
| Delegate | Hand off work to domain or specialist agents |
| Execute | Perform approved actions |
| Communicate | Send messages, draft communications |
| Verify | Check that actions produced intended outcomes |
| Reconcile | Compare intended vs. actual, feed to memory |
| Report | Surface findings to the user |
OODA Grammar
The Twin follows the OODA loop for every work item:
- Observe — Gather signals from the Spine
- Orient — Understand context and implications
- Decide — Propose a course of action
- Act — Execute under governance
The Four Buttons map to OODA:
| Button | OODA Phase | Action |
|---|---|---|
| Store in Spine | Observe | Capture truth |
| Ask your Twin | Orient | Understand the work |
| Assign your Twin | Decide | Delegate intent |
| Approve Twin's Action | Act | Authorize execution |
Suggested Actions Flow
Understand Context → Collect Evidence → Analyze → Recommend → Plan → Awaiting Approval → Execute → Reconcile → Complete
At "Awaiting Approval," the Twin pauses. The human must explicitly approve. The Twin never skips this gate.
Memory
The Twin maintains persistent memory:
| Type | Description |
|---|---|
| Episodic | What happened in specific situations |
| Semantic | Facts, relationships, patterns |
| Procedural | How to do things, workflows |
| User Preferences | How this user likes to work |
Memory is scoped to the user's role domain and governed by the same policies as the Spine.
Twin Rules
| Rule | Description |
|---|---|
| One Twin per user | Each person has their own Twin |
| Role-scoped | Twin's knowledge matches the user's role |
| Silent by default | Prefers silence over low-value commentary |
| Never acts without approval | Proposes, human decides |
| Operates on projections | Never sees unrestricted canonical Spine |
Twin Workbench Connects to
| Layer | Relationship |
|---|---|
| Spine Fabric | Reads via Projections, writes decisions |
| Projections | Receives sandbox context |
| Human Workbench | Shares work through Shared Task Layer |
| Coworker | Receives context from ambient companion |
| Work Allocation | Allocates and delegates work |
| Governance | Submits actions for approval |
| Agents | Delegates to specialist agents |
| Execution | Executes approved actions |
| Reconciliation | Verifies outcomes |