Part 1 of 4 · 1 min read
Overview
Research is the agent task where being fluent and wrong is the main risk: the output looks authoritative whatever the evidence. A good research agent is defined less by how well it writes than by how it shows its evidence, how it handles disagreement, and when it says "I could not find this".
About the example, and what is real
Real: the TF-IDF retriever, planner rules, sanitiser, conflict resolution, citation verifier and the 18 tests in code/mini-research; every number comes from running them. Invented: the four products, the 16 pages, their dates, trust levels and values, and the claim format (CLAIM entity | attribute | value) that lets a program check citations. Stand-ins: the rule-based planner (an LLM would plan) and the careless writer (random wrong citations and values); no model, no web, no real documents.