What it does
Every deployed agent eventually meets a situation it shouldn't handle alone — a regulated decision, an upset customer, a request that crosses the authority boundary it was given. Escalation to Human is the skill that recognizes those moments and hands off cleanly.
The hand-off is not a transfer button. It's a full briefing: the agent summarizes the conversation, names the trigger, identifies the right recipient (on-call rotation, named owner, role), and waits for acknowledgement before stepping back.
Triggers it watches for
- Authority boundary breached (the agent doesn't have permission to act on this)
- Emotional escalation (frustration, distress, anger above a configurable threshold)
- High-stakes decision (above a deal value, contract scope, or risk class)
- Explicit customer request to talk to a human
- Repeated failure to resolve (the agent has tried, and isn't getting closer)
When to use it
- Always — every Onyx deployment should have this skill enabled
- Pynk deployments where the agent acts on the user's behalf with third parties
How it works
The skill maintains an escalation profile per deployment: who handles what, business hours, on-call rotation, escalation chain. When a trigger fires, the agent generates a structured hand-off — situation, context, recommendation, urgency — and routes it to the right recipient through the configured channel.