Pattern: Escalation

Category: Human-in-the-Loop Source: FOR-0012 Status: Documented

When to Use

When an agent encounters a situation it cannot handle confidently — edge cases, ambiguous inputs, high-risk decisions, or errors it cannot self-correct. Rather than guessing or failing silently, the agent escalates to a human or a more capable system.

How It Works

• The agent assesses confidence in its ability to handle a situation
• If confidence is below a threshold (or the situation matches known escalation triggers), it escalates
• Escalation includes: what happened, what was attempted, why escalation is needed, and suggested next steps
• The escalation target (human, senior agent, or fallback system) receives the context and takes over
• Resolution is logged and optionally fed back to improve future handling
• The agent may continue with other tasks while waiting for escalation resolution

Example

A digital talent handling customer support for a small business encounters a request involving a potential legal issue (e.g., "I want to sue you for the defective product"). The agent recognizes this exceeds its scope, flags it as high-risk, and escalates to the business owner with full conversation context and a note: "This may require legal consultation — I have not responded to the customer."

Tradeoffs

Pro	Con
Prevents agents from making dangerous mistakes	Adds latency for escalated cases
Builds appropriate trust boundaries	Escalation triggers must be well-calibrated
Captures edge cases for future improvement	Over-escalation wastes human time
Clear accountability — agent knows its limits	Under-escalation risks undetected errors

Factory Usage

• Role Factory score gating: Roles scoring below 10/15 are returned to DESIGN rather than proceeding — an escalation from automated improvement to human redesign.
• Agent non-trigger redirection: When an agent recognizes a request is outside its scope, it redirects to the correct specialist — a form of peer escalation.