Skip to main content Instructions define how your agent behaves, what it knows, and how it should respond to users. Treat them as a living playbook that sets the agent’s personality, expertise, and operating rules.
What are Instructions?
Detailed prompts that guide behaviour and tone.
A place to declare expertise, escalation paths, or compliance requirements.
Guardrails that spell out what the agent should or shouldn’t do.
Inside the instructions editor, type @ to open the list of available tools and MCP resources.
Click the tool you want to reference—it is embedded directly into the prompt.
Describe when the agent should call that tool so usage is unambiguous.
Repeat for each tool you want the agent to understand. Embedded references stay synced if you rename the underlying tool.
Embedding tools in context eliminates guesswork; the agent knows exactly which capability to reach for and under which circumstances.
Best practices Writing effective instructions
Be explicit about the agent’s role, expertise, and tone.
Provide concrete examples of the behaviour you expect.
Set hard boundaries for topics or actions that are off limits.
Use the @ menu to anchor instructions to specific tools or MCP resources.
Document why a tool exists and the scenarios it supports.
Test the API or integration outside of Agent Builder before wiring it in.
Choose meaningful names and slugs so @mentions stay readable.
Configure authentication up front—agents can only call tools that succeed end-to-end.
MCP integration
Connect reliable MCP servers with predictable latency and uptime.
Note the context each endpoint returns so teammates know when to enable it.
Monitor responses over time and adjust endpoints or filters if the returned data drifts.
