Reading a research paper once for the whole lab

Data, ML and researchResearch group using agents to read and compare papers

Our lab of six kept asking agents to summarise the same papers. One 40-page paper on retrieval-augmented training was read by four people's agents in a month, each costing about 60k tokens, and each summary missed different details.

How can our lab stop paying for AI agents to summarise the same paper again and again?

Have the first reader's agent check ArcRoom's find_task with the paper's title or arXiv ID, claim it, then save a structured summary with complete_task: method, data, results and limits. Later readers get that summary and ask only follow-up questions. There is no freshness check for a PDF outside git, so if a v2 comes out, someone marks the old summary wrong.

Also asked as: Four people in our lab had agents read the same 40-page paper. How do we share one summary? · How do I share AI paper summaries across a research group? · Can AI agents reuse a summary of a paper someone already read?

The work

The first reader's agent checks find_task with the paper's title or arXiv ID, finds nothing, and claims it.

ArcRoom keeps
  • Reused answer
  • Claim
  • Saved answer
The next agent gets

find_task → match: reuse (0.93), saved by lena (Oct 3)

Step by step

  1. The first reader's agent checks find_task with the paper's title or arXiv ID, finds nothing, and claims it.
  2. It reads the paper and saves a structured summary with complete_task: method, data, results, limits.
  3. ArcRoom records the paper's URL, so the summary is tied to the source.
  4. Later readers get the saved summary and ask only follow-up questions.

What the agent sees

find_task → match: reuse (0.93), saved by lena (Oct 3)
"Paper: retrieval during pre-training. Method: kNN over 2T tokens. Result: +4.1 on QA at 7B. Limit: no eval below 1B, index cost not reported."

Compared with today's tools

Zotero notes or a shared Notion page

Whoever reads a paper writes notes in the shared library, and others check there before asking an agent.

Setup
30 min
Ongoing
write notes by hand, ~15 min per paper

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min per person (npx arcroom init)
Ongoing
None

A reference manager is better for citations and PDFs; ArcRoom catches the repeat read inside the agent, without anyone remembering to check.

Where it stops

There is no freshness check for a PDF outside git. If a v2 of the paper comes out, someone has to mark the old summary wrong.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in data, ml and research