Running multiple AI tools in parallel on a research task — one searching literature, one analyzing data, one drafting text — sounds appealing. The reality is messier. Here’s what actually works and what creates more work than it saves.
What Multi-Agent Means in Practice
A multi-agent research setup might look like: Claude handles writing and reasoning tasks, Elicit handles literature extraction, Cursor handles code, NotebookLM handles cross-paper synthesis. These tools run sequentially (one feeds the next) rather than truly in parallel, unless you build automation between them.
The Sequential Workflow That Works
A productive three-stage workflow: Stage 1 (Discovery) — ResearchRabbit maps the citation network, Semantic Scholar identifies the key papers, Elicit extracts structured data. Output: a Zotero library + extracted data table. Stage 2 (Synthesis) — NotebookLM synthesizes across the imported PDFs, Claude structures the argument. Output: annotated outline + draft synthesis. Stage 3 (Writing) — Claude drafts, you revise, Cursor handles any analysis code. Output: manuscript draft.
Where It Breaks Down
The handoff points are where time disappears. Exporting from Elicit, reformatting for NotebookLM, verifying that the AI synthesis is accurate — each handoff takes 30–60 minutes and requires your attention. The total time savings over doing it manually often disappoints until you’ve done the same workflow 3–4 times and streamlined the handoffs.
Automation Options
For researchers comfortable with code, Python automation between tools recovers real time. Zotero’s API lets you query your library programmatically. The Semantic Scholar API returns paper data in JSON. Scripting the data extraction from Elicit (CSV export + pandas) removes manual reformatting. Once built, these automations run in minutes rather than hours.
The Minimalist Alternative
Most PhD students don’t need a five-tool workflow. Claude + Zotero + one discovery tool (Semantic Scholar or ResearchRabbit) covers 90% of research support tasks. Add tools when a specific bottleneck appears — don’t add them preemptively because they seem useful in principle.
