Field noteNov 16, 20265 min read1 source

Google Cloud AI Research’s PaperOrchestra is an open multi-agent framework that turns research ideas and

In one minute

Google Cloud AI Research’s PaperOrchestra is an open multi-agent framework that turns research ideas and experimental logs into LaTeX manuscripts, including literature synthesis and generated visuals. Its value for builders is the concrete workflow: specialized agents handle outlining, review, writing, refinement and plotting, while users supply materials, a conference template and model credentials.

What changed: Google Research’s PaperOrchestra adds an automated multi-agent workflow that turns research ideas and experiment logs into submission-ready LaTeX papers.

What happened

Google Cloud AI Research presents PaperOrchestra as a framework for automated AI research paper writing. The repository describes a pipeline that transforms pre-writing materials—including ideas and experimental logs—into submission-ready LaTeX manuscripts. It says the output can include literature synthesis and generated visuals such as plots and conceptual diagrams. The named authors are Yiwen Song, Yale Song, Tomas Pfister and Jinsung Yoon, affiliated with Google Cloud AI Research. The repository includes a citation for a 2026 arXiv preprint, arXiv:2604.05018.

The workflow is divided among specialized agents. An Outline agent produces the structure; a Literature Review agent searches for, prioritizes and integrates relevant papers; Section Writing agents draft parts of the manuscript; Content Refinement agents improve the draft; and a Plotting agent can generate figures. The project also automates LaTeX output for a target conference template. The repository describes this as an end-to-end process, but its own setup still requires users to provide the research materials and template.

To run it, users can create a Conda environment with Python 3.11 and install dependencies from requirements.txt. A .env file holds credentials. The README names Semantic Scholar, OpenAI, Vertex AI and Gemini API variables, noting that users need the relevant keys for whichever model they want to use. The command-line interface takes a required raw-materials directory and a LaTeX-template directory, plus an output directory; if no output directory is specified, it creates a timestamped one. Plotting is optional and defaults to false. Users can configure separate model names for writing and literature agents, and reflection agents. There is also a bash script and an interactive Streamlit demo, with separate instructions for the latter.

The repository includes templates for target conferences, including cvpr2025, and says users can add templates in new subdirectories. It does not include the dataset used for paper generation; the README says it will be released separately later. The project carries a disclaimer that it is not an officially supported Google product. The repository does not state a price for using the code; model and API costs depend on the services a user configures.

What the archive adds

New Runtime’s earlier coverage described Google’s HappyDelivery working prototype, a multi-agent conversational-commerce system built with LangGraph, Gemini and Google Cloud (Building Conversational Commerce with Multi-Agent Swarms on GCP: A Working Prototype, 2026-10-28). PaperOrchestra is a new event, not a continuation of that commerce system: it applies agent orchestration to a different task, research-paper production.

Trend signal

This adds another example of work being divided among specialized agents rather than handed to one general-purpose agent. That is directionally consistent with AI-native organizations moving toward review and orchestration, though the repository describes a research-writing tool, not evidence about organizational adoption or measured productivity.

Connections to other events

The earlier HappyDelivery prototype also used a multi-agent architecture, but for conversational commerce. PaperOrchestra’s distinct contribution is to apply staged orchestration to scholarly writing, with literature search, section drafting, refinement and figure generation as separate responsibilities.

Why it matters

For teams building research or knowledge-work agents, PaperOrchestra makes a useful design choice visible: represent a long task as stages with distinct responsibilities and inputs, instead of asking a single agent to produce a finished artifact in one pass. The repository’s separate settings for writing/literature and reflection models also make model choice part of the workflow design. Builders can inspect how these responsibilities are packaged in a runnable project, then adapt the pattern to reports or other document pipelines.

The boundary matters as much as the automation. The tool accepts the researcher’s raw materials and target template; it does not remove the need to judge whether the cited literature is relevant, whether claims follow from experiments, or whether a figure accurately represents results. That makes review an essential part of using the output, not an optional polish step. For teams, the useful experiment is to test each stage against known work and decide where human approval is required before relying on generated manuscripts.

Limits and uncertainty

The repository describes intended capabilities, but does not report measured accuracy, citation quality, time saved, cost per paper or results across a benchmark. It does not establish that generated manuscripts meet any conference’s standards or that literature coverage is comprehensive. The dataset used for generation is not included and is described as a future separate release, which limits independent reproduction of the stated workflow. The project’s disclaimer also means users should not treat it as an officially supported Google product. API use may incur provider charges, but the repository gives no total operating-cost estimate.

Skill to build now

Take one completed project with experiment logs and a chosen LaTeX template, then run the pipeline with plotting disabled first. Review its outline, literature synthesis and drafted sections against the source materials before enabling figure generation. This isolates whether the writing stages are useful before adding another generated artifact to validate.

Reader takeaway

PaperOrchestra is a runnable example of multi-agent research writing, not evidence that papers can be safely automated end to end. Its reusable lesson is to separate the work into stages—and to keep review tied to the underlying experiments and sources.

Evidence / sources

  1. 1github.com/google-research/paper-orchestra
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicGoogleExplore the google topic hub.
  2. 02topicGeminiExplore the gemini topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract