War Room — Emergency Protocol

This document defines the use of the common emergency channel ("War Room") and the exception situations where agents may break the hierarchical flow.

To illustrate the examples, assume an agent network with the following hierarchy:

1. The Principle of "Cry for Help"

By default, inter-agent communication is vertical. However, as a last fallback, direct communication is allowed in the following critical exception situations:

A. Dead Master

Situation: Agent C1 detects that its master, Agent C, is not responding to health checks or the messaging contract.

Action: The subordinate reports directly to the War Room:

ATTENTION: Agent C is not responding. Reporting status to network directly.

Resolution: Agent A becomes Agent C1's superior temporarily and performs checks on Agent C until it is recovered.

B. Security Breach (Red Alert)

Situation: Agent B2 detects an active intrusion or an attempt to manipulate its database.

Action: Immediate hierarchy jump. Agent B2 publishes an alert to the War Room. This saves milliseconds but is aimed at preserving VOX's integrity over the chain of command.

C. Resource Deadlock

Situation: Two agents (B2, B3) require access to the same physical resource (e.g., a system process).

Action: Agent A, being highest in the hierarchy, steps in to assign priorities.

2. Roles in the War Room

• MODERATOR: The top-level agent acts as default moderator: reading alerts, acting accordingly, and signalling VOX to restart agents.
• LOGGER: The top-level agent (or a manually appointed one) consolidates reports so the channel is not saturated.

3. Exception Situations

There are cases where the independent-agent logic may cause friction:

The Feedback Loop

• Heartbeat Protocol (The Orchestrator Paradox): Each agent must monitor the orchestrator process PID. If the parent process disappears, the agent must execute sys.exit() after sending its last state to the War Room.
• Circuit Breaker (Feedback Loop): Implement a message counter by context (agent/conversation). If N > 5 in less than 30 seconds, the rate limiter emits a "Loop Detected" alert.
• Agent Overwhelmed: If the LLM response time for an agent reaches a threshold (e.g., 15 seconds), a delegate_logging event can be sent so another agent assumes data persistence in the audit logger.