Literature Review

A literature review is a systematic survey of what's been published on a topic. This is the workflow Agent Bayes was built around — start broad, narrow systematically, capture landmark works and debates, and end up with a structured map you can actually write from.

Before you start

  • Curate your KB first. Retrieval quality dominates everything downstream. If your KB is missing the seminal papers, the review will be missing them too. Use Zotero to bring in your existing library, or upload PDFs directly.
  • Pick a scope. "Late Bronze Age collapse" is a workable topic; "the ancient world" is not. The agent does better when the first instruction has a defensible boundary.

Step 1 — Establish the landscape

Start with one broad instruction. Don't pre-structure the mindmap; let the agent map out the territory.

Build an overview of the main scholarly perspectives on the Late Bronze Age collapse. Include the leading proponents of each view and the kinds of evidence they rely on.

The agent runs broad discovery. You should expect, at the end of this first workflow:

  • A root branch per major perspective or theme.
  • Each branch holding citation-backed claims summarizing that view.
  • Contradicting positions as sibling nodes — that's the system working as intended.
  • A textual summary describing what it covered and any gaps it noticed.

If the breadth feels off (too narrow, too wide), correct in the next instruction. Don't try to get it perfect in one shot.

Step 2 — Audit the breadth

Before going deep, scan the top level of the map:

  • Are all the major schools represented?
  • Are the citations weighted reasonably (no one paper dominating)?
  • Did the agent flag KB limitations? If so, that often means a foundational text isn't in your library — add it before going further.

This is the cheapest moment to add sources. Every later step gets better if the corpus is right now.

Step 3 — Go deep, one branch at a time

Select a branch and instruct the agent to expand it. Selection scopes the agent's context to that subtree.

Expand the systems-collapse view. Cover the supporting evidence, the main critics, and any recent revisions to the position.

Repeat for each branch you care about. Going branch-by-branch is significantly better than one mega-instruction — the agent's context stays focused, retrieval is more on-target, and you can review after each pass instead of after everything.

Step 4 — Surface debates and contradictions

Once branches are populated, ask the agent to find friction:

Across the current mindmap, identify the points where these perspectives directly contradict each other. Create sibling-node groupings under shared parent topics where the disagreement is sharp.

This restructures the map so disagreements are visible at a glance, instead of buried inside two unrelated branches.

Step 5 — Find what's missing

Use the Gap Finding workflow as a follow-up: ask the agent to identify under-explored areas, weakly-supported claims, or open questions visible in the current map. This is where a review stops being a summary and starts being a contribution.

Step 6 — Restructure for writing

When the content is settled, ask the agent to reorganize the map into a draft outline (see Outlining a Paper). The Editor handles this without re-running retrieval — fast and cheap.

Tips that pay off

  • Don't dismiss low-confidence claims. Inspect them. Low confidence often points to genuinely contested ground, which is exactly what a review should cover.
  • Use labels like primary-source, review-paper, weak-methodology early. They make Step 4 and Step 5 much easier. See Labeled Items.
  • Save your RAG queries for the searches you find yourself running by hand. The next person reviewing this topic (often: future you) will thank you.

What's next