INTENT Layer

Understanding What Agents Want to Do

The first layer of BASIS governance — before enforcement, before logging, understand the intent.

---

What is INTENT?

The INTENT layer is responsible for:

1. Parsing — What action is the agent attempting?
2. Planning — What steps are required?
3. Risk Surfacing — What could go wrong?
4. Structuring — Format for downstream processing

┌─────────────────────────────────────────────────────────────┐
│                        INTENT LAYER                         │
└─────────────────────────────────────────────────────────────┘

     ┌─────────────────┐
     │  Agent Request  │
     │  "Send email    │
     │   to client"    │
     └────────┬────────┘
              │
              ▼
     ┌─────────────────┐
     │     PARSE       │──▶ Action: send_email
     │                 │    Target: external
     │                 │    Data: client_contact
     └────────┬────────┘
              │
              ▼
     ┌─────────────────┐
     │      PLAN       │──▶ Step 1: Lookup client
     │                 │    Step 2: Compose email
     │                 │    Step 3: Send via SMTP
     └────────┬────────┘
              │
              ▼
     ┌─────────────────┐
     │  RISK SURFACE   │──▶ Risk: MEDIUM
     │                 │    - External communication
     │                 │    - Contains client data
     │                 │    - Irreversible
     └────────┬────────┘
              │
              ▼
        STRUCTURED INTENT
              │
              │ Passes to ENFORCE layer
              ▼

---

Why INTENT First?

You can't enforce what you don't understand.

Without INTENT	With INTENT
"Agent did something"	"Agent attempted send_email to external recipient"
Binary allow/deny	Risk-appropriate response
Black box decisions	Explainable governance
Post-hoc auditing	Pre-execution understanding

---

INTENT Output Schema

interface Intent {
  // Identification
  intentId: string;
  agentId: string;
  sessionId: string;
  timestamp: string; // ISO8601
  
  // What
  action: string;
  parameters: Record<string, any>;
  
  // Plan
  plan: {
    steps: Step[];
    estimatedDuration: number;
    reversible: boolean;
  };
  
  // Risk Assessment
  risk: {
    level: "minimal" | "limited" | "significant" | "high";
    factors: string[];
    mitigations: string[];
  };
  
  // Requirements
  capabilities: string[];
  resources: string[];
}

---

Risk Classification

INTENT is responsible for initial risk assessment:

Minimal Risk

• Read-only operations
• Public data access
• Internal computations
• No external effects

Limited Risk

• User data read/write (scoped)
• Internal communications
• Reversible operations

Significant Risk

• External communications
• Sensitive data access
• System modifications
• Multi-step operations

High Risk

• Financial transactions
• Bulk data operations
• Irreversible actions
• Permission changes

---

API Endpoint

POST /v1/intent/evaluate

Request:

{
  "agentId": "ag_7x8k2mN3p",
  "intent": {
    "action": "send_email",
    "parameters": {
      "to": "client@example.com",
      "subject": "Invoice #1234",
      "body": "..."
    }
  },
  "context": {
    "userId": "usr_abc123",
    "sessionId": "ses_def456"
  }
}

Response:

{
  "intentId": "int_9h8g7f6e",
  "status": "evaluated",
  "plan": {
    "steps": [
      {"action": "validate_recipient", "risk": "low"},
      {"action": "compose_email", "risk": "low"},
      {"action": "send_email", "risk": "medium"}
    ]
  },
  "risk": {
    "level": "medium",
    "factors": [
      "external_communication",
      "contains_financial_data"
    ]
  },
  "capabilities": [
    "communication/send_external",
    "data/read_user"
  ],
  "nextStep": "enforce"
}

---

Implementation Requirements

Requirement	Description
REQ-INT-001	Generate unique intentId for every evaluation
REQ-INT-002	Identify all required capabilities
REQ-INT-003	Classify risk level (4 levels)
REQ-INT-004	Structure output per schema
REQ-INT-005	Complete evaluation in < 500ms
REQ-INT-006	Never execute actions (evaluation only)

---

Next Layer

Once INTENT has structured the request, it passes to ENFORCE for trust verification and policy checking.

[INTENT] ──structured intent──▶ [ENFORCE]