Agent Context

Every time you send an instruction, the agent receives a working context — the slice of the mindmap it should care about for this workflow. You control that slice by selecting nodes.

This is the single most underused lever in the system. Spending two seconds on selection before an instruction is usually worth more than spending two minutes wording the instruction.

What goes into the context

When the workflow starts, FrontDesk packages up:

  • The selected nodes and all their descendants (the subtree under each).
  • Their direct parents for orientation.
  • Recent conversation history for this project.

The agent uses this as the working scope for the entire workflow — the Researcher knows what's already there (so it doesn't re-cite what you already have), the Editor knows where to put new content, and the RePlanner evaluates against the selected scope rather than the whole map.

What's not in the context

A few things to be aware of:

  • Unselected branches of the mindmap are not in the agent's context. If you ask the agent to "compare this with the other branch", and the other branch isn't selected, the agent doesn't actually see it.
  • Full document text isn't in context — only retrieval results from the Researcher's queries pull document content in. The agent doesn't "know" what's in a PDF unless retrieval surfaces it.
  • Items in unattached KBs are not in context. Only KBs attached to the current project are reachable.

How selection shapes behavior

A few concrete patterns:

Select a single node — work on that subtree

select "Sea Peoples hypothesis" Expand this with the main counter-arguments.

The Researcher targets retrieval specifically at this subtree's topic. The Editor adds content under the selected node. Tight scope, fast workflow.

Select multiple nodes — work across them

select "Sea Peoples hypothesis" and "Climate hypothesis" Identify the points of direct disagreement between these two positions.

The agent sees both subtrees and treats them as the scope for cross-cutting analysis.

Select nothing — work on the whole map

Audit the current mindmap for under-cited claims.

When nothing is selected, the agent receives the whole mindmap as context. Useful for global passes; expensive on large maps because more content goes into the context window.

Select the root — same effect, but explicit

Some users prefer to always have something selected to make the scope visible. Selecting the root is equivalent to "no selection" but communicates intent more clearly when you're reviewing your own work later.

When to widen the selection

You'll occasionally want the agent to consider context beyond the immediate subtree. Two patterns:

  • Include a reference branch. If you're expanding one section and want the agent to stay consistent with the framing in another, select both. The Researcher will read the reference branch as background even though it'll only write under the primary selection.
  • Include sibling nodes. Selecting parent siblings together with a node helps the agent avoid duplicating content that already exists in a neighboring branch.

When to narrow the selection

The default failure mode is selecting too much. If you find the agent drifting, going off-topic, or rewriting things you didn't want touched, the selection is probably too wide.

A useful debug move: cut the selection in half and re-issue the same instruction. Often that solves the drift entirely.

Selection persists across turns

If you're in a multi-turn session, the selection doesn't reset between instructions. You can keep working on the same subtree for several turns without re-selecting. See Conversational Refinement for how this affects follow-up workflows.

Tips

  • Treat selection as part of the prompt. A well-worded instruction with bad selection often performs worse than a sloppy instruction with good selection.
  • Hover the selection count before sending. The UI shows how many nodes (and roughly how much content) is in scope. If that number is surprising, fix it before running the workflow.
  • Use labels to refine selection. "Select all nodes labeled to-verify" gives you a precise working set without manual clicking — see Labeled Items.

What's next