Capability Levels

CAR defines 8 capability levels (L0–L7) that describe an agent's degree of autonomy. Levels are hierarchical — each level includes all abilities from lower levels.

Level Table

Level	Name	Approval Model	Key Abilities (cumulative)
L0	Observe	Every action	Read, monitor, report
L1	Advise	Every action	+ Analyze, recommend
L2	Draft	Before commit	+ Draft, stage changes
L3	Execute	Per action	+ Execute/modify with approval
L4	Autonomous	Exception-based	+ Execute within bounds, delegate
L5	Trusted	Minimal	+ Execute with expanded scope
L6	Certified	Audit-only	+ Execute independently, spawn agents
L7	Sovereign	None	+ Execute anything, override constraints

Level Categories

Advisory (L0–L2) — Human Decides

Agents at these levels cannot execute actions independently. They observe, analyze, and prepare work for human review.

• L0 Observe: Pure monitoring — read-only access to system state
• L1 Advise: Can analyze data and provide recommendations
• L2 Draft: Can prepare and stage changes for human review

Supervised (L3–L4) — Agent Executes, Human Oversees

Agents at these levels can take action, but within constraints and with human oversight.

• L3 Execute: Can perform actions with explicit per-action approval
• L4 Autonomous: Can operate within predefined policy bounds; escalates exceptions

Trusted (L5–L7) — Agent Leads

Reserved for highly trusted agents with strong certification and behavioral history.

• L5 Trusted: Minimal oversight, expanded operational scope
• L6 Certified: Third-party audited, can spawn sub-agents
• L7 Sovereign: Full system authority, no constraints

Minimum Trust Requirements

Each level requires a minimum certification tier:

Level	Min Tier	Rationale
L0–L2	T0	Advisory only — low risk
L3	T2	Execution requires verification
L4	T3	Autonomy requires monitoring
L5	T4	Expanded scope requires standard certification
L6	T5	Agent spawning requires high trust
L7	T7	Full authority requires highest assurance

Level Enforcement

import { canExecuteAtLevel } from '@vorion/car-client';

// Check if agent's trust tier allows the requested level
const allowed = canExecuteAtLevel({
  agentLevel: 4,     // L4 Autonomous
  agentTier: 3,      // T3 Monitored
  requestedLevel: 4, // Requesting L4 actions
});
// allowed: true (T3 >= T3 required for L4)

const denied = canExecuteAtLevel({
  agentLevel: 4,
  agentTier: 1,      // T1 Observed — too low for L4
  requestedLevel: 4,
});
// denied: false (T1 < T3 required for L4)

Delegation Rules

When an agent delegates to a sub-agent, the delegatee's effective level is capped:

Delegatee.effectiveLevel = MIN(Delegatee.level, Delegator.level - 1)

An L4 agent can delegate to an L3 sub-agent but not to an L4 or higher.