Capability Fabric

Capability Fabric is the governed execution language between intelligence and operational action. It answers the question:

WHAT CAN BE DONE?

What Capability Fabric Does

• Discovers capabilities from the registry
• Selects capabilities by policy and context
• Validates parameters against capability contracts
• Enforces governance gates before execution
• Routes execution to the correct runtime implementation
• Returns the result as canonical evidence

What Capability Fabric Does NOT Do

• Own canonical truth
• Decide outcomes alone
• Bypass governance
• Provide direct operational memory

INTELLIGENCE
      ↓
CAPABILITY SELECTION
      ↓
POLICY / GOVERNANCE
      ↓
EXECUTION
      ↓
RESULT
      ↓
CANONICAL CONTINUITY

Capability Registry

The capability registry is the catalog of all available capabilities. Each capability is defined by:

Field	Description
key	Unique identifier (e.g., retry_payment, create_task)
plane	Which architectural plane it operates in
inputs	Expected parameters and types
outputs	Return values and types
required_scopes	Authorization scopes needed
governance_policy	What governance is required
surface_compatibility	Which surfaces can invoke this capability

Governance Contract

Every capability execution path must include a governed boundary:

Capability selection
      ↓
Policy evaluation
      ↓
Governance gate if required by policy
      ↓
Execution contract accepted
      ↓
Capability execution
      ↓
Canonical result path

Rule: Capability Fabric must not bypass governance.

Implementation Types

Capability implementation can be:

Type	Description
API	Direct REST/GraphQL call to external service
MCP	Tool invocation via Model Context Protocol
Spine Tool	Internal Spine operation
Workflow	Multi-step orchestrated flow
Agent	AI agent execution
ADK	Agent Development Kit mechanism
Provider	External provider mechanism

Discovery and Marketplace

Three related marketplaces support capability discovery:

Marketplace	Purpose
Integration Marketplace	External system connectors
Plugin Marketplace	Workbench extensions
MCP Marketplace	AI tool capabilities

Lifecycle:

Discover → Validate → Install/Connect → Authorize → Register Capability → Govern → Execute

Rules:

• Marketplace is capability discovery/distribution
• Marketplace is not canonical truth
• Plugin installed ≠ permission
• MCP available ≠ authority

Capability Examples

Capability	Description	Governance
retry_payment	Retry a failed payment	Approval required
create_task	Create a new work item	Policy check
send_customer_message	Send message to customer	Approval required
update_account	Update account record	Role check
create_invoice	Generate an invoice	Approval required
schedule_meeting	Schedule a meeting	Policy check
reconcile_payment	Match payment to invoice	Role check

Capability and Execution

Capability is separate from execution. The flow is:

CAPABILITY (what can be done)
      ↓
EXECUTION FABRIC (how it is done)
      ↓
GOVERNANCE (may it be done)
      ↓
RESULT (what happened)
      ↓
RECONCILIATION (is it correct)

Related

• Execution Fabric — how capabilities are executed
• Governance — the policy boundary
• Twin — who invokes capabilities
• Agent Fabric — the agent fleet that uses capabilities