This is where an agent takes shape. You give it a name, a purpose and the boundaries that keep it on track. Agents are configured under Admin Tools > Settings > AI Agents. Before configuring your first agent, set your organization context under Settings on the same page. That context applies to every agent you build, so it's worth doing once up front. See Setting Your Organization Context for details. Agent Settings To add a new Agent, select ti ti-robot Agents then click ti ti-square-plus. Name - Make this descriptive. "Staff Chat Assistant" is more useful than "Agent 1."Type - Chat or MCP. See Type section. Audience - Internal or Public. See the Audience section. Description - Internal notes only. This doesn't affect how the agent behaves. Use it to record which team or purpose the agent serves.Instructions - How the agent should act. See the Instructions section. Avatar - Optional. Upload an image to give the agent a visual identity in the chat panel.Role - Right now "Default" is your only option. In the future the Role field will be used for AI model selection.Auto Summarize Threshold - How many tokens into a conversation before the agent starts summarizing earlier messages to stay focused.Exclude System Skills - Check this to prevent Rock's built-in system skills from being available to this agent.Current Person Template - A Lava template that tells the agent who is currently logged in. Leave this blank unless you have a specific reason to customize it. Type Chat agents live inside Rock. Staff access them through the docked chat panel, which you an access from anywhere in Rock by clicking the ti ti-sparkles agent button. This is the right starting point for most organizations. MCP agents connect Rock to an external AI tool running on someone's computer, like Claude Desktop. Once connected, that tool can look up records and take actions in Rock using the person's own security permissions. See Set Up an MCP Agent for the full walkthrough. Audience Audience is one of the most important decisions you'll make for each agent. Internal agents are for staff and trusted volunteers. They work with your full data set and the full range of skills you assign.Public agents are for anyone: visitors to your website, people filling out a form, anyone who might interact without a staff login. The skills you attach to a public agent should reflect that. Keep them narrow and purpose-built. If a skill surfaces personal data you wouldn't hand to a stranger, it doesn't belong on a public agent. Picking this setting helps the tools adjust themselves knowing that they are talking to a public audience. Instructions Instructions are where you shape how the agent behaves. Think of it as a short onboarding note: tell the agent how to act, what to prioritize and what to avoid. A few things worth including: How to handle vague requests ("Ask one clarifying question before taking any action that changes data.")Tone ("Be warm and concise. Write like a colleague, not a formal report.")Boundaries ("Never delete a record without explicit confirmation.")Specific context needed by your Agent Keep it short. A long list of rules is harder to follow consistently than a clear set of priorities. Start simple and add only when you notice a real gap. Only add what's needed. Testing Before sharing an agent with your team, test it yourself, more than once. AI doesn't produce identical results every run, so test your most important prompts at least a few times and look for consistency. Also test as the people who will use it. Log in as a volunteer and try it. If the agent surfaces something that person shouldn't see, the fix is in your Rock security, not the agent configuration.