3.4 Agent System
An Agent System is a container for building an AI assistant that answers questions from data in the Data Context Hub. You build it on a drag-and-drop canvas by placing one or more Agents, attaching Memories that give those agents access to data, and connecting agents together when you need a multi-step workflow. You then run and test the system through the built-in Chatbot.
An Agent System combines three kinds of building block:
- Agents: The AI components that process a query. Each agent runs on a model defined by a Provider Configuration and follows its own instructions.
- Memories: Data connections that give agents access to specific entities, projections, or files. A memory is attached to an agent on the canvas.
- Chatbot: The interface for running the system, asking questions, and inspecting how each agent responded.
The canvas
The canvas is a drag-and-drop editor. You drag an Agent onto it to add one, open an agent to configure it, and attach a Memory to give it data. To build a multi-step workflow, you draw a connection from one agent's handle to another so the first agent's output feeds the second. Memories connect to agents through a separate handle. See Agents for how to configure and connect agents, and Memories for the memory types.
Execution order and agent groups
You do not assign a role to an agent. When the system runs, it works out the execution order from how the agents are connected: agents with no incoming connection run first, and each remaining agent runs once all the agents feeding into it have finished. A connection that forms a cycle is rejected, because there would be no valid order.
The system info panel, and the Chatbot run Overview, group the agents by their place in that order into three sections:
- Intake: The entry agents (no incoming connection) that receive the query and pull data from their memories.
- Process: Any agents in between that refine or transform the intermediate results.
- Output: The final agents whose response is shown to the user.
A single-agent system is the common case, and that one agent is simply both intake and output.
Versioning
An Agent System is versioned so you can iterate safely. Each version has a state:
- Draft: A work-in-progress version you can edit and test.
- Published: A version promoted for production use. Published systems appear under the Published tab of the Agent Systems list; everything else is under Unpublished.
- Archived: A retired version, kept for history.
The Versions panel in an Agent System shows the timeline of versions with their state and dates. Select a version to review it. Publish the selected version with Publish version, or remove it with Delete version. Editing a version that is already published automatically creates a new draft, so a published version is never changed in place.
Create and manage an Agent System
- Open the Agent Systems section from the sidebar and click Add new.
- Enter a title and a short description, then Save.
- Open the system to build it on the canvas: add Agents, attach Memories, and connect agents as needed.
- Click Run system to open the Chatbot and test it.
From the list you can Search and filter systems, switch between the Published and Unpublished tabs, Duplicate system to copy one for reuse or testing, and Delete system to remove one.
