Conversational Refinement

Research isn't a one-shot prompt. The interesting work happens across many turns — propose, react, narrow, branch. Agent Bayes preserves the full conversation per project, so follow-ups read naturally and the agent stays aware of what you've already decided.

The conversation is real

Each project carries one continuous conversation with the agent. Every instruction you send and every summary the agent returns is stored and visible to the agent on subsequent turns. This is what makes references like "expand that section", "the previous result", or "the second branch you just added" work — they're not magic, they're just pronouns the agent can resolve.

Follow-up patterns that work

Reactive refinement

(previous workflow expanded the Sea Peoples subtree)

That's too speculative. Re-run with stricter confidence requirements, and drop any claim that rests on a single source.

The agent reads the previous result, the relevant nodes are already there, and the new instruction adjusts the work. No need to restate the topic.

Incremental narrowing

Add the main critics of the systems-collapse view.

(workflow runs, results land)

Of those critics, the post-2018 ones are the most relevant. Drop the older ones, and expand each remaining critic with their specific counter-evidence.

Two turns, each one strictly more focused than the last. This is how a tight, defensible section gets built.

Branching mid-conversation

Looking at the Sea Peoples subtree, this is making me wonder whether the Egyptian inscriptions are being read tendentiously. Build a parallel subtree on that question.

The agent treats the new direction as an extension, not a reset. Existing context is preserved; new work happens alongside.

Course correction

Wait — go back. The restructure you just did dropped the "primary evidence" sub-nodes I wanted to keep. Restore those and redo the restructure with them preserved.

The conversation is the audit trail. The agent can refer back to what was done and undo or adjust accordingly. You can also browse version history for a hard revert.

Interrupting and resuming

Long workflows (especially the full 3-loop refinement cycle) can take 90–180 seconds. You can:

  • Let it run — leave the tab open, work elsewhere, come back when the summary lands.
  • Interrupt mid-flight — stop the workflow. Any nodes already applied stay; the rest is rolled back. The summary explains what was completed.
  • Resume — issue a follow-up that picks up where you stopped. Workflows checkpoint automatically, so you can pause complex investigations, gather additional materials, and continue from the last completed evaluation cycle.

Interruption is non-destructive. Partially completed work is committed to the map only when the Editor has finished its current edit batch; otherwise it's discarded.

Selection across turns

Selection persists between turns unless you change it. That's usually what you want — you're often working on the same subtree for several instructions in a row. Just keep an eye on it: if you switch topics mid-conversation, re-select before sending the next instruction or the agent will work on the wrong scope. See Agent Context.

When to start over

Sometimes the right move isn't to refine but to discard a workflow's output entirely:

  • The original instruction was confused or wrong.
  • The result drifted in a direction you don't want to engage with.
  • The branch is a dead end.

In those cases:

  • Revert to the previous mindmap version (clean) — see Version History.
  • Re-issue with a clearer instruction.

This is cheaper than trying to refine your way out of a bad starting point.

When the agent should ask you a question

The agent doesn't currently pause to ask for clarification mid-workflow — it makes a best-guess call and continues, then reports what it decided. If a decision was wrong, that's your cue for a follow-up.

If you anticipate ambiguity, pre-empt it in the instruction: "If you're unsure whether to merge or branch, branch."

Tips

  • Keep instructions short within a conversation. You don't need to repeat context the agent already has. "Now do the same for the climate branch" is plenty.
  • Use conversational references freely. "That last citation", "the third sibling you added", "the previous result" all work.
  • Switching projects switches conversations. Each project has its own history; there's no cross-project memory.

What's next