Agent layers
Global prompt
The always-on layer shaping identity, tone and guardrails on every turn. It runs regardless of which
playbook or workflow is active, and is written in four default sections: #Role, #Goal, #Tone
and #Guardrails.
Also asked as:
- the instructions that apply to every single reply
- set the bot’s personality and house rules once
- the system prompt for the whole assistant
Related: Instructions · Guardrails · System prompt
Documented in: Global prompt › What is the global prompt? · Global prompt › Best practices · Global prompt › #Role · Global prompt › #Goal · Global prompt › #Tone
Instructions
The routing layer: which playbook or workflow handles what. Where the global prompt defines how the agent behaves, instructions define what it does with a request.
Also asked as:
- tell the assistant which skill handles which kind of request
- the decision layer that picks a playbook
- where the agent’s routing rules live
Related: Global prompt · Trigger · Reference
Documented in: Instructions › What are agent instructions? · Instructions › Writing agent instructions
Start message
The first thing the agent says. An agentic start message is written under a # Starting Message
heading in the instructions; a fixed greeting is a Message step in the initialization workflow with
Wait for user input turned on.
Also asked as:
- the greeting the bot sends before the visitor types anything
- open every chat with the same welcome line
- make the assistant speak first
Related: Instructions · Initialization workflow
Documented in: Instructions › Adding a start message · Initialization workflow › Static start message
Trigger
What tells the agent when a playbook or workflow applies. Older material calls this the “LLM Description”; the current name is Trigger. System tools carry one too, describing what the tool does and when to use it.
Also asked as:
- describe when a skill should be used
- how does the model know which playbook to pick
- the description the router reads
Related: Instructions · Playbook · System tools
Documented in: Playbooks › Creating playbooks · Querying the knowledge base › Trigger
Behaviour
The model, voice and memory settings an agent runs under: default model, memory, timezone, voice output and input, session timeouts, chat persistence, events and the global no-match and no-reply responses.
Also asked as:
- the settings page for the model, the voice and memory
- where do I change what the agent runs on
- agent-wide settings rather than per-playbook ones
Related: Default model · Memory · Voice output
Documented in: Behaviour
Agent tools
The tools attached to the agent itself rather than to one playbook. Also called global tools: APIs, integrations, MCP servers and functions added at the agent level, so every playbook can use them without adding each one individually. They apply in agentic contexts only: a workflow calls tools through its own steps.
Also asked as:
- make one tool available in every playbook
- attach a tool at the agent level
- share a tool across all skills
Related: Tool · System tools · Playbook
Documented in: Global tools › Adding a global tool · Global tools › When are global tools available
Global tools
The tools added at the agent level — APIs, integrations, MCP servers and functions — so every playbook can use them without adding each one individually. The same thing as Agent tools; the Studio page is titled Global tools.
Also asked as:
- add a tool once for every playbook
- the agent-level tools list
Related: Agent tools · Tool
Documented in: Global tools › Adding a global tool
Guardrails
Guardrails
Non-negotiable rules the agent must always follow, written under a # Guardrails heading in the
global prompt. Models are tuned to pay extra attention to that section. Good guardrails are specific
and actionable; “be professional” is a tone instruction, not a guardrail.
Also asked as:
- rules the bot must never break
- hard limits on what the assistant may say or do
- where do I put the things the agent is not allowed to do
Related: Global prompt · Preventing hallucination · Staying in scope · Transaction safety
Documented in: Global prompt › #Guardrails
Preventing hallucination
A guardrail pattern: the agent says it does not have the information instead of guessing, never combines partial knowledge base results with assumptions, and never invents a plausible reply when a tool call fails.
Also asked as:
- the assistant invents facts that are nowhere in our content
- keep replies to what the knowledge base actually says
- the bot confidently gives wrong information
- the assistant invents answers our documents never gave
- stop the assistant from guessing
- only answer from the documents we uploaded
Related: Guardrails · Hallucination · Grounding
Documented in: Global prompt › #Guardrails
Protecting sensitive data
A guardrail pattern: never reveal internal pricing logic, margins or cost structures, never share one customer’s information with another, never read back full card numbers, SSNs or passwords.
Also asked as:
- stop the assistant from leaking private details
- make sure the bot never reads out a card number
- keep internal pricing out of replies
Related: Guardrails · PII redaction
Documented in: Global prompt › #Guardrails
Staying in scope
A guardrail pattern that keeps the agent to its domain: no legal, medical or financial advice, no comments on competitors, and an offer to connect with the right team when the conversation drifts outside it.
Also asked as:
- keep the assistant on topic
- refuse questions unrelated to our product
- the bot answers things it has no business answering
- the assistant answers questions that have nothing to do with our business
- keep the bot from going off topic
- refuse unrelated requests politely
Related: Guardrails · Global no-match
Documented in: Global prompt › #Guardrails
Transaction safety
A guardrail pattern for actions with consequences: no refund, cancellation or account change without explicit customer confirmation, business rules never overridden, identity verified before account data is touched.
Also asked as:
- make the bot ask before it does anything irreversible
- require an explicit yes before a refund goes through
- verify who the customer is before changing their account
- the bot confirmed a refund without double-checking
- ask the customer to confirm before acting
Related: Guardrails · Initialization workflow
Documented in: Global prompt › #Guardrails
Writing a strong global prompt
The practices the docs recommend: keep it short (100–1000 words; every word runs on every turn), be direct (declarative rules, not hedging), keep it static so the model can reuse its cache, and keep step-by-step procedures out of it — those belong in instructions, a playbook or a workflow.
Also asked as:
- how long should the global prompt be
- my prompt is huge and replies got slower
- what belongs in the global prompt versus a playbook
Related: Global prompt · Instructions · Playbook
Documented in: Global prompt › Keep it short · Global prompt › Be direct · Global prompt › Variable Consistency · Global prompt › Common mistakes · Global prompt › How long should my global prompt be?
Agent runtime
Combines instructions, tools, knowledge, and conversation state to decide how to respond.
Documented in: Agent runtime