Part 7 of 11 · 2 min read
Answering research-agent questions in an interview
How would you build a research agent you can trust?
"Plan the facts needed, search per need within a budget, and read pages as untrusted data. Rank sources by kind and date and show conflicts, write each claim with a source, and verify each citation against the page before the report goes out. It says 'not found' when evidence is missing. I evaluate it on questions with gold answers and traps, measuring coverage, correctness, citation precision and cost." Follow-up: "How do you stop hallucinated citations?" A verifier that checks the cited page contains the claim, with regeneration from read evidence on failure. In the example's toy experiment it cut bad citations from 4.6 to 0.
Two sources disagree. What does the agent do?
"Prefer the primary and more recent source, but show both values with sources and dates, and flag the answer if only low-trust sources exist. If it cannot decide, it asks the user or says the sources conflict." Follow-up: "What if both are secondary?" Look for the primary origin; otherwise report the disagreement and the reasons to prefer one.
How do you defend a research agent against prompt injection on web pages?
"Treat page text as data, drop instruction-like sentences, mark and downgrade suspect pages, keep the research phase read-only with no sensitive tools, and never let page content change the plan or the output format. I add planted-instruction pages to the eval set." Follow-up: "Is that enough?" No; filters miss variants, so the permission boundary (least privilege) is the real defence (Level 10 modules 1 and 2).
How do you control the cost and latency of deep research?
"Budgets on sources and steps, cheaper models for reading and a stronger one for synthesis, caching fetched pages, parallel fetches, early stopping per need, and durable execution so a failure does not restart everything. I report cost per answered question." Follow-up: "When is deep research not worth it?" When a single retrieval answers the question; deep research trades high cost and latency for depth.