AI Agents

Rock has a heavy load, holding all the data that your organization runs on, and we know that you have a heavy load too. Rock has always been about empowering staff and simplifying processes so people can focus on ministry.

Agents are the next step. Think of them as digital helpers that free you up for ministry or for that task you've been putting off for months.

As a developer building agents and the tools that power them, you carry real responsibility. Agents must stay safe and easy to understand. The Rock Core team has worked hard to ship a powerful yet safe set of starter Skills. If you plan to add to that list, here is what we've learned and what we use to keep agents and tools both powerful and safe.

Terminology

There are a couple of terms you should know that will be used throughout this guide.

Agent

An Agent can be thought of as both an AI configuration set as well as a persona.

Agents allow you to define various configuration elements including what skills and tools are available for use when interacting with the agent. You can also set security on an agent to dictate who can and can't use the agent. Skills sometimes have configuration options, which are defined when you add the skill to an agent.

The persona is also defined on an agent. This covers things like how verbose or concise you want the responses to be. How polite it should be when crafting responses.

Skill

You can think of a Skill as a functionality group. For example, there is a Finance skill. This skill contains all the functionality for working with financial accounts, donations, etc. Skills sometimes provide instructions for how and when the tools should be used by the agent. Skills can also be secured so even if they exist on an agent, the person using the agent must have access to the skill.

A skill is made up of one or more tools.

Tool

Tools provide functionality for an agent to "do" something. It might be to lookup a person in the database, add somebody to a group, or tell you what campuses are available.

A tool is, generally speaking, a single unit of work. If you ask the agent to "update Ted Decker's birth date to Aug 3, 1984 and add him to Alisha Marble's small group" that will probably trigger 5 tool calls.

1. Lookup Ted Decker
2. Update Ted Decker's Birth Date
3. Lookup Alisha Marble
4. Find a Small Group led by Alisha Marble (by Id)
5. Add Ted Decker to Small Group (by Id)

Tools can also be secured, so a person must have access to a tool before they are allowed to use or even see it.

Example

Each skill is a collection of related tools. The Group Skill contains many tools for working with groups. When you assign a skill to an agent, all of its tools come with it. You can view, edit and secure individual tools from the skill detail page.

Context Window

You can imagine the context window as the cognitive brain capacity of the agent. As humans, we can only hold so many things in our mind while working on them. When discussing a TV show with a friend, we mentally pause when trying to remember a specific scene. This is because we don't walk around with everything we have ever seen or read in our mental context.

For AI agents the same rule applies. They have access to absurd amounts of data, but they can't hold it all for active processing.

The context window is made up of a number of things, but the two biggest ones are the system prompt and the chat history. The system prompt is static, but the chat history grows longer the more you chat. Eventually the context window gets so full that the agent must do one of two things. Stop talking to you, or summarize the discussion to shrink the context. As you can guess, they choose the latter. But this comes at a price, that summary is just that: a summary. Many of the specific details are lost.

Therefore, it is critical when designing tools to keep in mind how much data you are adding to the context. If it is useful data, by all means include it. But if it isn't really useful, exclude it. Or make it optional so it is only included if they specifically ask for it.

System Prompt

The system prompt contains instructions for the agent. These are often directives that must be followed. For example, one directive might be "Don't call any "delete" tools if there is any ambiguity in what the user asked you to do." After all, we don't want the wrong thing deleted just because the request was vague. The system prompt is made up of various individual pieces that, altogether, become the system prompt. You can read more about this here.