Undermind

Undermind positions itself as the research agent for people who actually read papers, not just skim abstracts. The autonomy here is in how it searches: you give it a question, it reads hundreds of abstracts, refines its query based on what it finds, and follows citation trails without you babysitting each step. That matters when you're doing a proper literature review and don't want to miss the seminal paper buried under different terminology. I tested it on a niche cardiology question about microvascular dysfunction biomarkers. Standard PubMed searches gave me the usual suspects. Undermind found three older papers I'd missed because they used different terms, and it surfaced a 2019 review that became the backbone of my reading list. The citation tracing is the standout feature: it doesn't just find papers that cite your seed paper, it evaluates which citations are actually relevant to your specific question. That's harder than it sounds and saves hours of dead ends. The interface is clean. You write a question in natural language, it shows you its search strategy, and you can steer it if it's wandering. The free tier gives you a handful of deep searches per month, which is enough to evaluate it but not enough for a systematic review. Paid plans unlock more searches and faster results. Failure modes: it's only as good as what's indexed, so grey literature and preprints are hit-and-miss. It also struggles with very new topics where citation graphs haven't formed yet. And if your question is too broad, it drowns you in results without much filtering. Compared to Elicit, Undermind is narrower but deeper. Elicit is better for quick evidence synthesis across many papers. Undermind is better when you need to be confident you haven't missed a key paper in a specific domain. Consensus sits somewhere in between but leans more towards answering yes/no questions than building a reading list. The research monitoring feature emails you when new papers match your saved searches. I haven't used it long enough to judge if the signal-to-noise ratio holds up, but the idea is sound for anyone tracking a moving field.