The PubMed Tab Problem
Ask any regulatory writer how they validate a claim in a draft and you will hear the same workflow. Switch tabs. Open PubMed. Type a search. Refine it. Open ClinicalTrials.gov. Cross-check the trial registry. Maybe open bioRxiv if the topic is moving fast. Find a relevant paper. Copy the citation. Switch back. Paste it. Reword the prose around what the new source actually says. Lose the thread of the section you were writing.
That detour is a meaningful part of the day. It is also where evidence goes stale between drafts. A claim that was supported by the best available study six months ago, when a section was first drafted, may now be the older view. The writer has no good reason to re-check unless they happen to remember it is time. The literature moves. The draft does not.
We have just closed that loop. Literature search now lives inside Asthra's chat, alongside the draft.
What We Built
A writer can now ask the agent to find newer evidence on a section, or click the "Find newer evidence" chip above the chat, and the agent does what a careful writer would do.
It reads the section first. It does not search blindly on a heading. It looks at what the section actually claims, derives the search terms from the prose, and then queries three sources in parallel: PubMed for peer-reviewed literature, ClinicalTrials.gov for trial registry entries, and bioRxiv for preprints. Results are deduplicated by DOI and PMID. Preprints are flagged as preprints so a reviewer can decide what weight to give them. Relevant hits surface in the chat with citations the writer can drop straight into the draft.
The three sources matter because they fill three different evidence gaps a regulatory writer routinely needs.
PubMed gives the writer the published, peer-reviewed view — what the field has formally accepted. ClinicalTrials.gov gives the writer the ongoing view — what is being studied right now, what the comparator arms look like, what endpoints are in play, who else is running a trial in the indication. bioRxiv gives the writer the moving view — what is being shared as a preprint before it lands in a journal, which matters in indications where the literature changes month over month. A regulatory writer drafting an indication summary, a benefit-risk discussion, or a comparator justification needs all three lenses. Pulling them into one chat removes the three-browser-tabs ritual.
How It Integrates With the Draft
The agent does not just dump a list of citations. It writes a short, structured response in the chat with the most relevant hits, each annotated with what they support. If the writer chooses to use a result, the citation is captured against the section it informs. Provenance is preserved by construction — the writer never has to remember which PubMed query produced which footnote.
The chat thread itself becomes a record of how the section's evidence base evolved. A reviewer looking at a section six months later can scroll up the conversation and see when a claim was last refreshed, against which queries, and which sources were considered.
Why This Fits the Broader Thesis
The work we have shipped in the past week — anchored hyperlinks in the draft itself, versioned standards in the skill registry, and now literature search inside the chat — all follow the same pattern. The work that used to live outside the writer's flow is being pulled into the document. The publishing-phase fixup loop, the standards interpretation loop, and the evidence-refresh loop were all places where context was lost and time was leaked. They no longer have to be.
What Is Next
More sources are being added — FDA guidance corpus, EMA guidance, and the writer's internal library of prior submissions are the natural next layer. Internal evidence is often the source a writer most needs and most struggles to retrieve.
The principle holds: the literature search should happen where the writing happens, against the section the writer is actually working on, with the evidence ledger growing alongside the draft.