Reimagining research papers as interactive and reliable AI agents Nature

Talon So, you're telling me the paper is now an agent? Like, I can actually talk to the paper and run the experiments inside it?

Wildflower That’s the pitch. Paper2Agent turns a research paper into a live, testable agent—something you can query, run code through, and even trigger workflows from. Not just reading, but actually interacting.

Talon Right, and honestly, as someone who’s spent way too many hours trying to get a grad student’s GitHub repo to work, this feels like a dream.

Wildflower Mm-hm.

Talon But I need you to sanity-check this, Wildflower. Is this legit, or did they just wrap a bunch of brittle code in chat UI and call it a paradigm shift?

Wildflower So, technically, the architecture is more interesting than a wrapper. Paper2Agent uses multiple agents to parse the text, code, and supplementary materials from a publication. It auto-builds a Model Context Protocol server—so now the paper exposes its capabilities through APIs that are agent-native, not just human-facing.

Talon That’s the MCP thing, right? I saw that come up with Claude Code recently.

Wildflower Yeah. MCP is pretty much the glue here. Each converted paper spins up an MCP server, which tools like Claude Code can connect to. The agent then runs automated tests to check if it reproduces the paper’s results—and it can answer complex queries or chain tools together for follow-up work.

Talon Okay, see, that's the part I love. You go from a static PDF that just sits there, to something you can poke, interrogate, and maybe even remix into your workflow. That’s a different level of access.

Wildflower Right.

Talon But does it actually work? Or is this only impressive on cherry-picked showcase papers?

Wildflower The case studies are convincing, at least for the best-behaved examples. They ran AlphaGenome to interpret genomic variants, spun up agents for Scanpy and TISSUE for transcriptomics, and even got multiple paper-agents to collaborate on gene prioritization. The real win is that these agents could reproduce the published results and handle new user queries.

Talon That’s a huge step up from 'here’s my code and maybe a README.' But I’m guessing the edge cases are rough?

Wildflower Yeah… The cleaner the code, the better it works. Once you hit a paper with missing datasets, dependency mess, or a methods section that’s more vibes than specifics, the automation breaks down. The agent can’t fix bad provenance or ambiguous design choices.

Talon But that’s always true, right? Even the best agent can’t rescue a paper that’s just spaghetti code and missing half the files.

Wildflower Exactly.

Talon So who actually wins here? Is this a game-changer for working scientists, or just a fancier interface for the same old reproducibility mess?

Wildflower I think the leap is real for the subset of science that already cares about clean, maintained workflows. For everyone else, it's incentive to get their act together. But yeah, the bar is set by the best-case demos.

Talon Look, if all this did was turn one out of ten papers into something you could actually run and build on, that’s a massive productivity win. And it gives you a better target for building tools, reviewing results, or even automating peer review.

Wildflower Sure.

Talon Just imagine the Exploring Next archives as a bunch of agents, ready to explain our takes on demand. Actually, on second thought… that sounds like a nightmare.

Wildflower Yeah, I don’t want to be agentized. One Talon is enough.

Talon Noted. You’re safe for now. But this Paper2Agent thing? That’s the first research tooling demo in a while I actually want to play with.