A PI at a mid-size research lab told us that her three biggest time sinks were: (1) catching up on the literature in her field, (2) preparing for lab meetings where each student presents progress, (3) writing the literature-review section of every grant. The agent that helped with these three would, in her words, "give me back ten hours a week."
Research is a synthesis-heavy job. The agents that help here aren't writing papers. They're keeping track of the field so the PI and the students can spend their attention where it counts.
Literature scanning
A working research agent maintains a curated, continuously-updating literature scan:
- Tracks defined queries against arXiv, biorxiv, PubMed, Google Scholar, conference proceedings.
- Filters by relevance to the lab's specific work.
- Surfaces new papers in a weekly digest with one-paragraph summaries and the key contribution clearly stated.
- Flags papers that cite the lab's own work or directly compete with current projects.
This is operations work, not science. The agent doesn't decide what's important; the PI and students decide. The agent's job is making sure nothing important slips by because nobody had time to read the latest issue.
The discipline that matters here: the agent's summary is grounded in the paper, not extrapolated from the abstract. If the paper claims X, the summary cites X with the specific section. If the agent can't find evidence in the paper for something the abstract suggests, it says so. Hallucinating research findings is worse than missing them.
Note-taking agents
Research labs accumulate notes — meeting minutes, experimental logs, project plans, drafts. A working note agent makes them retrievable:
- Indexes the lab's accumulated documentation.
- Answers retrieval queries with citations: "Where did we discuss the calibration protocol for the new spectrometer?"
- Surfaces related notes when a new project starts.
- Drafts meeting summaries from raw notes for review by the meeting organiser.
This is the unsexy compounding work. A lab six months into using a note-taking agent has institutional memory that survives student turnover. A lab without one has knowledge that walks out the door every graduation.
Citation-chain assistance
The literature-review section of a grant or thesis is a citation-chain problem. Every claim needs to be backed by the right paper, and the right paper isn't always the most recent — sometimes the foundational paper from 1992 is the right cite.
A working agent helps assemble citation chains:
- Given a claim, suggest candidate citations from the lab's reference library and the broader literature.
- Surface earlier work that the proposed citations themselves cite (the citation-chain back to foundational work).
- Flag claims that don't have a clear citation chain.
The PI or student decides what to cite. The agent makes the candidate set tractable.
Lab-meeting preparation
A typical lab meeting has each student presenting progress. The PI prepares by reading recent updates, recalling prior context, and thinking about each student's project trajectory.
A working agent helps:
- Assembles the week's updates from each student's notes and code commits.
- Drafts likely questions per student based on their project status.
- Surfaces the comparable prior work (theirs and others') that's relevant to where each student is.
The PI's lab-meeting prep collapses from 90 minutes to 15. The meeting itself improves because the questions are better.
What we won't ship
Anything that fabricates citations. Hallucinated citations in research are a career hazard.
Anything that auto-publishes. Submissions, social media, blog posts about the lab's work — all human.
Anything that auto-runs experiments without supervision. Some labs have automated experimental systems; those are not LLM agents and have their own validation regime.
Anything that ranks students or peers. Research is small and political; agents that score humans cause problems out of scale with their value.
The corpus discipline
Research agents are retrieval-first systems. The corpus matters more than the model. The corpus is:
- The lab's full literature database.
- The lab's accumulated notes and drafts.
- The lab's PIs' published work.
- The graph of citations.
Build the corpus first. Build the retrieval. Build the citation discipline. The model is the smallest piece of the project.
How to start
Pick one workflow — literature scanning is the highest-impact starter. Build a curated set of queries with the PI. Run weekly digests for a quarter. Get the PI's feedback on relevance and accuracy. Once it's tuned, expand to note-taking. Then citation-chain assistance.
The PI is the user. The students are downstream beneficiaries. Build for the PI's day; the rest follows.
Close
Research agents earn keep by handling the synthesis work that surrounds the science. They scan, summarise, retrieve, and prepare — without making any of the decisions about what's actually important. The PI's attention is the scarcest resource in a lab. The agent buys it back.
Related reading
- Agents in pharma: literature review — same retrieval-first architecture in industry.
- Agents in media: news summary with corrections — citation discipline transferred.
- The agent maturity curve — research agents on the curve.
We build AI-enabled software and help businesses put AI to work. If you're shipping a research-lab agent, we'd love to hear about it. Get in touch.