November 18, 2026Open SourceAgentsResearch

PaperOrchestra: Google Open-Sources the Paper-Writing Agent Team

Notes and experiment logs in, a submission-ready LaTeX paper out. That is the pitch of PaperOrchestra, a multi-agent pipeline from Google Cloud AI Research, and on Nov 17 it picked up 1,175 GitHub stars in a day, roughly 40% of its 2,873 total. https://github.com/google-research/paper-orchestra

To be clear about dates: the repo was created on May 1, so this is not a new project. What changed is that it became usable. The latest commits, on Nov 15, are community fixes for missing dependencies, optional API keys and LaTeX escaping, and then the trending wave hit. It is Apache-2.0, runs on Gemini or OpenAI backends, and is by Yiwen Song, Yale Song, Tomas Pfister and Jinsung Yoon.

The pipeline is a writing team with roles. One agent builds the outline. One does the literature review through Semantic Scholar. Others write sections, refine, and draw plots and diagrams. The input is deliberately messy, the kind of raw notes a researcher actually has, rather than a tidy draft. The evaluation dataset is not included yet and is promised later, so the quality claims cannot be checked from the repo alone.

This is useful and a little alarming at the same time. Writing up results is the least loved part of research, and a tool that turns logs into a first draft saves real weeks. But reviewers are already drowning, and the cost of producing a plausible paper just dropped to near zero while the cost of checking one did not. The interesting question is not whether agents can write papers. It is who builds the agent that reads them critically.
← Previous
Claude Code Team: The Test Suite Is the Real Product
Next β†’
Three Papers, One Lesson: Tools Lie and Agents Believe Them
← Back to all articles

Comments

Loading...
>_