Skills that change the design

Skills add capabilities to a raia cX agent. Most are configuration details, but a few change the shape of the whole solution. This page covers those. For every skill and its settings, see the Skills directory.

Escalation

Escalation is a decision made during a conversation. It can be triggered by specific keywords or by broader criteria you describe in a prompt. You configure the response, who it's assigned to, whether a person takes over, and who is notified by email, SMS or webhook. Escalated conversations appear in Copilot.

If escalation only needs to alert someone, the agent can keep helping. If risk or policy requires a person, switch the conversation away from automated handling.

Use escalation for in-the-moment decisions, such as a customer asking for a person. Use the end-of-conversation webhook when you need a complete record of the whole conversation.

Learn more: Escalation.

Scoring, summaries and notifications

When a conversation is complete, it can be summarized and scored, and a notification can be sent based on timing or a score threshold. The score can follow whatever rubric you write, but it must match the use case. Support urgency and sales interest are different measures.

Write the rubric down, with examples, exceptions and what each threshold triggers. Don't treat a general sentiment score as a priority rule unless the business owner has agreed to it.

Learn more: Scoring and Notifications.

Orchestrator

With the orchestrator, a primary agent asks specialist agents for help, gathers their answers and replies to the person, without handing the conversation over. Routing can follow keywords or general instructions, and you control how much of the conversation each specialist sees. Specialists use their own knowledge and tools.

Good reasons to use it:

An internal support agent consults a public-knowledge specialist, without customers ever reaching internal knowledge.
A light front-line agent calls a stronger specialist only when a record needs to be created and checked.
A general product agent routes to separate experts for distinct products.

Use the orchestrator when there's a real difference in data access, expertise, model or quality control. Don't split one simple job across agents for appearance's sake. Share only the context each specialist needs.

A call to a specialist is a one-off request, not a conversation. The specialist answers that request; it doesn't remember earlier exchanges unless you pass them in.

Learn more: Orchestrator.

MCP: two directions

The MCP skill works in two directions that are easy to confuse:

Direction	What it does	Use it when
Agent as an MCP server	Makes your raia agent callable from Claude, ChatGPT and other MCP clients	People should reach the agent from their own AI assistant
Agent calling MCP servers	Gives your raia agent tools in external systems	The agent needs to read or act in another system

When a provider offers a public MCP endpoint with simple token sign-in, connect it directly. When sign-in is more involved, or you're wrapping an existing API, use a bridge managed in Command.

Exposing an agent as an MCP server also enables a useful cost pattern: a light agent retrieves and filters information, then hands a compact, sourced answer to a stronger model. See Cost and model choice.

Learn more: MCP and MCP integrations.

Webhooks

Webhooks fire on conversation events, such as a conversation starting or ending, a status changing, or a manual trigger. The end-of-conversation event can include the final summary and score.

A common shape is: a conversation event fires a webhook, which starts a Loop in Command, which asks an agent to do the work, waits for approval if needed, and takes the action.

Anything receiving a webhook should check the request is genuine, handle duplicate deliveries safely, validate the content, and protect the conversation data inside it.

Learn more: Webhooks.

Agent settings that matter

Setting	Turn it on when
File search	The agent should answer from its trained knowledge. Usually on.
Code interpreter	The agent works with numbers, counts, spreadsheets or tables. Don't rely on the model for arithmetic.
Structured (JSON) output	Another system needs consistent, machine-readable results. Not for ordinary chat.
Shared vector store	A test agent should reuse another agent's knowledge. This also shares access, so check the audiences match.
Default error message	Always. Decide what the agent says when data is missing or a tool fails.

Use reusable packs for common guardrails rather than rewriting them on every agent. Instructions are versioned, but a separate test agent is still the safest way to try changes.

Learn more: Launch Pad and agent setup and LLM models.