Pre-processor

The Pre-processor runs before the main agent sees a user message. Use it to enforce safety policies, classify intent, or apply deterministic rules — without changing the main agent’s logic.

​

Sentinels

Sentinels are mini-agents that evaluate the incoming message using an LLM. They can block, tag, or transform the request based on their classification. Select one or more Sentinels from those configured in Settings → Sentinels. Execution order: Sentinels run sequentially in the order they’re listed. If a sentinel blocks a request, subsequent sentinels and the main agent are skipped. The blocking sentinel’s configured response is returned to the user. Common sentinel use cases:

• Block off-topic or harmful inputs before they reach the agent
• Classify intent and tag the request for downstream logic
• Detect language and route accordingly

​

Empathy Rules

Empathy Rules are condition-based rules that run without an LLM call — they’re evaluated deterministically based on patterns, keywords, or metadata. Much faster and cheaper than sentinels. Each rule consists of:

• Condition — what to match (keyword in message, metadata value, user attribute)
• Action — what to do (respond with a fixed message, tag the request, escalate)

Example rules:

• IF message contains “cancel my account” → respond with “Please contact our retention team at…”
• IF message contains profanity → block with “Please keep the conversation respectful”
• IF user_language = “es” → tag as spanish and forward

​

Combining sentinels and empathy rules

Both can be active simultaneously. The execution order is:

1. Empathy Rules (instant, no LLM)
2. Sentinels (LLM-powered, sequential)
3. Main agent

This ordering minimizes cost — deterministic rules catch the easy cases before any LLM is invoked.

​

Next steps