Google’s New AI Agents: One Draws Your Figures, the Other Reviews Your Paper

Google Research just dropped two new AI agents aimed at the academic workflow, and honestly, they’re tackling two of the most painful parts of the research process: making decent figures and dealing with peer review.

If you’ve ever tried to get a methodology diagram past a reviewer at a top-tier conference, you know the pain. You spend hours in Illustrator or Inkscape, tweaking arrow styles and color palettes, only to have someone ask you to change everything. Meanwhile, the peer review system itself is buckling under the sheer volume of submissions. Reviewers are overworked, evaluations are inconsistent, and everyone feels it.

PaperVizAgent (formerly called PaperBanana, which I think is a better name) is a multi-agent system that generates publication-ready figures from text. You give it two things: your method section and a detailed figure caption. It then orchestrates five specialized agents—a retriever, a planner, a stylist, a visualizer, and a critic—to produce the final illustration. The critic agent loops back with feedback if something doesn’t match the source text, which is a nice touch.

I’ve seen automated figure generation before, and it usually produces something that looks like a confused toddler drew it with crayons. But Google’s results here are genuinely impressive. They claim it outperforms GPT-Image-1.5, which is a high bar. The examples in their paper show complex methodology diagrams with proper flow arrows, consistent color schemes, and readable labels. For statistical plots, it generates executable Python code, which is a sane approach—you can tweak the code if you don’t like the output.

Then there’s ScholarPeer, the automated reviewer agent. This one is more controversial, but also potentially more impactful. It takes an academic paper and produces a review that’s grounded in the literature. It doesn’t just check formatting; it evaluates the methodology, checks citations, and provides critical feedback. Google says it beats state-of-the-art automated reviewers, which isn’t saying much given how bad most automated reviews are, but the examples look surprisingly thoughtful.

The obvious concern here is that we’re automating the human judgment that’s supposed to ensure scientific quality. But let’s be real: the current system is already broken. Reviewers are tired, biased, and inconsistent. ScholarPeer could at least provide a baseline review that’s thorough and consistent, then let human reviewers focus on the nuanced parts. Or it could be used for pre-screening, flagging papers that need more work before they even hit a human’s desk.

I have mixed feelings about this. On one hand, anything that reduces the administrative overhead of academic research is welcome. On the other hand, I worry about a future where papers are written by AI, reviewed by AI, and the only humans involved are the ones reading the final product. But we’re not there yet, and for now, these tools seem genuinely useful.

The code for PaperVizAgent is already available, which is good. ScholarPeer’s paper is out too. I’d love to see how these perform in the wild, especially against real reviewer feedback. The bar is low, but the potential is high.