Asking follow-up questions

When a run finishes, the conversation is not over. You can keep asking the agent questions about the sources it gathered — drilling into a mechanism, comparing two studies, or chasing a number — and every answer stays anchored to the papers already in your run, with citations you can open.

#Continuing a run with follow-ups

Every run leaves you with an evidence base: the papers the agent retrieved from the life-science literature, the passages it read, and the synthesis it produced. A follow-up is a new question posed against that same evidence base. You are not starting a fresh search and you are not talking to a general-purpose chatbot — you are continuing a conversation about your run and the specific sources behind it.

The agent keeps the context of everything that came before, so you can build on earlier answers without restating the setup. Ask "What about in mouse models?" or "How does that compare to the second study?" and it resolves "that" and "the second study" from the thread rather than making you respell the whole question. Each turn adds to the run instead of replacing it, so the notebook and its citations grow as you explore.

When a follow-up needs detail that the abstract alone cannot supply, the agent reaches for the full text of the relevant papers — methods, results, and discussion — rather than guessing. Once a paper's full text has been pulled into the run, it is reused for later questions, so the deeper you go the more grounded the answers become.

#Kinds of questions you can ask

Follow-ups are most useful when they push on the evidence — asking the agent to summarize, contrast, extract, or reconcile what the sources actually say. Common shapes include:

• Summarize — condense the takeaways across the sources, e.g. "Summarize the key findings from these papers in three bullets."
• Compare — set studies side by side, e.g. "Compare the methodologies used across these trials" or "How do the two dosing regimens differ?"
• Extract a specific detail — pull a number or fact out of the text, e.g. "What p-values were reported for treatment efficacy?" or "What was the sample size in each study?"
• Explain a mechanism — walk through how something works, e.g. "Explain how this signaling pathway is regulated according to these papers."
• Surface contradictions — ask where the sources disagree, e.g. "Are there any conflicting findings among these studies?"
• Identify gaps — ask what is still open, e.g. "What limitations or research gaps do these papers flag?"

More specific questions produce more focused answers. "What was the median progression-free survival in the phase III study?" will land better than "Tell me about survival," because the agent can point at an exact passage rather than paraphrasing broadly.

#How answers stay grounded in sources

The defining property of a follow-up is that it is grounded. The agent answers only from the papers in your run, and every claim it makes is footnoted back to the source it came from. That gives you three things on every turn:

• Source attribution — each statement is tied to a specific paper, so you can see which study a claim rests on.
• Verification — you can open any cited paper and read the passage yourself, rather than taking the synthesis on faith.
• Traceability — the path from a sentence in the answer back to the source publication is explicit end to end.

Because answers are constrained to your run's sources, the agent has far less room to invent things. When the evidence in front of it does not actually answer your question, it says so — flagging that the selected papers are insufficient — instead of filling the gap with a confident guess. If a follow-up heads somewhere your current sources cannot support, that is usually a signal to widen the run rather than to trust an ungrounded answer.

#Example questions

A follow-up thread often starts broad and narrows as you learn what the sources contain. A typical progression through a single run might look like:

1. "Summarize the main conclusions across these papers."
2. "Which of those conclusions is supported by more than one study?"
3. "For the strongest result, what effect size and confidence interval were reported?"
4. "Did any of the studies contradict that finding?"
5. "What about in mouse models — do the sources cover preclinical work?"
6. "What limitations do the authors themselves acknowledge?"

Each turn builds on the last: the agent already knows what "that finding" and "the strongest result" refer to, so you can move quickly from a broad overview down to a single number and its caveats — with a citation attached at every step.