The Multi-Agent System

When you communicate with the agent, it's not handled by one model. Agent Bayes uses multiple specialized agents that handle your requests, to assure accuracy and efficiency. You don't need to think about them most of the time — but knowing roughly how they interpret your requests makes the system's behavior much less surprising.

Intent analysis and scope

Before any research or editing happens, the system interprets your message in the context of everything you have provided: the mindmap nodes you have selected or pinned as context, and the conversation history.

The same message — "expand this" — means something very different depending on whether you have a single node selected, an entire subtree pinned, or nothing focused at all. The scope you set determines what gets worked on.

If the system finds the provided instructions ambiguous, it may ask you a clarifying question before proceeding.

See Agent Context for how to control what context the agent sees.

How a workflow runs

Based on that intent analysis, each message takes one of three paths:

  • If the answer is already on the mindmap, it responds directly without searching your documents.
  • If your request is purely structural or stylistic — renaming a node, fixing grammar — it makes the edit without doing any retrieval.
  • If new evidence is needed, it searches your knowledge bases, synthesizes citation-backed claims, and writes them to the mindmap.

After each research pass, the system evaluates whether the request has been fully addressed. If meaningful gaps remain and they look retrievable, it runs another pass with a narrower focus. This can repeat up to three times before the workflow ends. You'll see nodes appear on the mindmap as they are written, not all at once at the end.

See also

What's next

  • Terminology Graph Explorer — a tool that helps you understand the relationships between concepts in your knowledge base, and how the agent uses them during research.