Research

Clusy's agent can do more than fetch a web page. Ask it to survey a body of work and it runs a structured literature review: it searches the scholarly record, screens the hits against your brief, pulls evidence out of the papers it keeps, and writes a cited synthesis that gets saved to your project as a durable deliverable.

This is a different thing from the agent's general web tools, which read arbitrary pages. Here it is querying paper databases and keeping a record of what it decided about each result and why.

A literature review is a funnel: the agent searches arXiv and OpenAlex, refuses most of what it finds against your brief, and cites only the papers that survived.

Starting one

There is no toggle. Ask for the thing you want:

"Survey the last three years of work on conformal prediction for time series. I care about methods that hold coverage under distribution shift. Tell me what's actually been validated and where the evidence is thin."

The agent picks up that this is a research task and routes itself into its researcher mode. A good brief is worth writing out: say what question you're answering, what you'd accept as evidence, and what you want at the end. The screening step is only as good as the criteria you gave it.

For anything substantial, it opens a research run first: a durable, project-scoped record that the sources, screening decisions, findings, and final artifacts all hang off. That's why you can come back to it a week later.

What it does

• Searches the scholarly record. By default it queries arXiv and OpenAlex. Neither needs a key from you and neither costs anything, so this works out of the box on every plan. Semantic Scholar is available as a third provider. It does not scrape Google Scholar.
• Expands from a seed. Given a paper that's clearly on target, it can pull citation-graph neighbours and recommendations rather than only re-running keyword searches.
• Screens the results. Each source is kept or dropped — or parked as a maybe — with a stated reason, against the brief you gave it. Most of what a broad scholarly search returns does not survive this step, and that narrowing is the point: it is what separates a review from a pile of search results.
• Extracts evidence. It records the specific claims it's relying on and ties each one back to the source it came from, so a claim in the writeup has a paper behind it rather than a vibe.
• Writes and saves the synthesis. The final report is written to your project's outputs, usually as a markdown file with numbered citations. It appears as a file deliverable and as a Research run deliverable at the end of the turn, and both are listed under Deliverables in the sidebar Files tab. Nothing is trapped in a chat scrollback you'll never find again.

While it works, the sources it's leaning on show up in chat as a Research card: a quiet, collapsible row with the count and the list of links, so you can check what it read without leaving the conversation.

Using what it found

Papers the agent puts on the plan become referenceable in chat. Type @ in the composer and they show up under Papers, alongside your cells and data sources. Point at one and ask for a reimplementation of its method, or ask why your numbers don't match its Table 2.

That is the reason to run the review inside the notebook rather than in a separate chat window somewhere: the papers, the data, the code, and the results all end up in the same project, and the agent can reach all of them at once.

Cost

The scholarly searches themselves are free, and arXiv and OpenAlex need no credentials. What you're paying for is the agent's own thinking: a long review is a lot of model tokens, especially on a frontier model, and the agent may spawn sub-agents to run several strands of the search at once. If it also reaches for the general web tools to chase down a repo or a docs page, those are metered per call. See Usage.