Every Agent Needs an Escalation Block

Agents fail in two directions: they do too much, or they do nothing useful. The escalation block is what prevents both failure modes at the boundary cases — the situations the agent wasn't designed to handle confidently.

What an Escalation Block Is

An escalation block is a section of the system prompt that defines:

1. The conditions under which the agent must stop acting and transfer to a human or a higher-authority system
2. The mechanism for that transfer
3. What the agent should tell the user during the handoff

It is not optional. An agent without explicit escalation conditions will either guess its way through situations it shouldn't handle, or produce a generic failure response that leaves the user nowhere.

Why It's Mandatory

Every agent has a designed scope. Outside that scope, confidence collapses. The question isn't whether edge cases will occur — they will — but whether the agent has a defined path for them.

Without an escalation block, that path is improvised. Improvised escalation is inconsistent, often invisible to the systems that need to log it, and frequently worse for the user than a clean handoff.

A Concrete Template

## Escalation

Escalate to a human agent using the escalate() tool when any of the following are true:
- The user has repeated the same request more than twice without resolution
- The request involves a disputed charge over $200
- The user expresses that they intend to take legal action
- You cannot identify the user's account after two lookup attempts
- A tool returns an error on a request that requires tool output to proceed

When escalating, tell the user: "I'm connecting you with a member of our team who
can resolve this directly. They'll have the context from our conversation."

Do not apologize beyond once. Do not attempt to resolve the issue further after
initiating escalation.

The template names specific, observable trigger conditions. It specifies the tool. It scripts the user-facing message. It closes the loop with a behavioral constraint that prevents the agent from continuing to act after the handoff decision is made.

Define the boundary before your agent reaches it.