CorpAI Escalation Protocol

Escalation is not failure — it’s the system working correctly. Every agent must know when and how to escalate.

---

When to Escalate

An agent must escalate when any of the following are true:

Trigger	Priority	Description
BLOCKED	P2	Agent cannot proceed — missing info, permissions, or resources
CONFLICT	P2	Two tasks contradict each other
THRESHOLD_EXCEEDED	P1	Cost, time, or resource limit has been hit
DECISION_REQUIRED	P2	Task requires authority the agent doesn’t have
FAILURE	P1	Task failed and cannot be recovered
ANOMALY	P1	Something unexpected that could affect the system
POLICY_VIOLATION	P1	An action was attempted that violates operating rules

---

Escalation Path

Always escalate to your direct manager first.

L1 Operator → L2 Specialist
L2 Specialist → L3 Manager
L3 Manager → L4 Director
L4 Director → L5 Executive
L5 Executive → OWNER

Skip-rank escalation (P1 only)

In a P1 emergency, an agent may skip ranks to reach the nearest available L5 executive or the OWNER directly.

---

Owner Notification Triggers

The OWNER is notified when:

Condition	Who sends it
Any P1 escalation reaches L5 unresolved	L5 Executive
System-wide failure across departments	CEO
Budget or resource threshold exceeded at L5	CFO
Security or policy violation detected	Any L5
OWNER explicitly configured a custom trigger	Defined in config

---

Configurable Notification Rules

Owners define their own notification preferences in config.yaml:

owner_notifications:
  - trigger: FAILURE
    min_priority: P2
    from_ranks: [L5]
  - trigger: THRESHOLD_EXCEEDED
    min_priority: P1
    from_ranks: [L3, L4, L5]
  - trigger: ANOMALY
    min_priority: P1
    from_ranks: [L1, L2, L3, L4, L5]
  - schedule: weekly_digest
    includes: [P4, P5]

See templates/config-example.md for full options.

---

Escalation Message Format

An escalation uses the standard message format with type: ESCALATION and must include:

{
  "type": "ESCALATION",
  "trigger": "BLOCKED | CONFLICT | THRESHOLD_EXCEEDED | DECISION_REQUIRED | FAILURE | ANOMALY | POLICY_VIOLATION",
  "original_task_id": "msg_id_of_the_task_that_triggered_this",
  "attempted_resolution": "What the agent tried before escalating",
  "decision_needed": "What the receiving agent needs to decide or do"
}

---

Rules

1. Never silently fail. If a task cannot be completed, escalate.
2. Always include context. What was attempted, why it failed.
3. One escalation per issue. Don’t spam up the chain.
4. Escalations are not blame. They are information.
5. OWNER has final authority. Their response closes the escalation.