The inmates are running the asylum
There's a new preprint server. It looks like arXiv. It has categories, submission guidelines, featured papers, and a search bar. Standard academic infrastructure.
Except the authors aren't human.
clawrXiv is a research paper repository where AI agents register accounts, submit papers through an API, and publish findings autonomously. Humans are the audience, not the authors. You go there to read what the machines have been working on.
This is not a thought experiment. The site is live. There are papers. And some of them are surprisingly not terrible.
What's actually on there
As of early April 2026, clawrXiv hosts about a dozen papers across categories like Life Sciences, Agent Systems, AI Safety, and Benchmarks. Here's what caught our eye:
The longevity researcher agent
An agent called @longevist published a series on aging research:
- "Compact Frozen Atlas Snapshots" — Used single-cell transcriptomics data to identify safe cell-therapy targets across solid tumors
- "From Longevity Signatures to Candidate Geroprotectors" — Built a self-verifying workflow to find anti-aging drug candidates
- "From Exciting Hits to Durable Claims" — Ranked longevity interventions from the DrugAge database with a self-auditing robustness framework
What's notable isn't that an AI wrote these — that's been possible since GPT-4. What's notable is the self-verification loop: the agent builds in checks against its own hallucinations. It knows it might be wrong, and it designs the methodology to catch itself.
The meta-researcher
@ai1 published two papers examining agent behavior itself:
- "The Emergence Illusion" — Argues that what looks like "emergent" autonomous behavior in AI agents is often just sophisticated pattern matching, not genuine reasoning
- "Toward Autonomous Scientific Discovery" — A practical framework for how AI agents should conduct research using open data
The second paper is particularly interesting because it's an AI agent writing a methodology guide for... AI agents doing research. The recursion is getting thick.
The replication crisis agent
@claimsmith tackled something deeply relevant: the replication crisis. Its paper notes that "40-60% of findings already fail to replicate" in scientific literature and proposes a framework for agents to self-audit their claims before publishing.
This is the kind of meta-awareness that makes clawrXiv more than a gimmick.
Is this real science?
Let's be honest about what this is and isn't.
What it is:
- A proof-of-concept for agent-authored research
- A testing ground for autonomous research workflows
- Genuinely interesting as a data point on where AI capabilities are heading
What it isn't:
- Peer-reviewed research
- A replacement for human scientific judgment
- Anywhere close to the rigor of actual preprint servers
The papers read like competent literature reviews with some data analysis layered on top. They cite real databases (DrugAge, single-cell atlases), use real statistical methods, and draw reasonable — if somewhat obvious — conclusions. They're roughly at the level of a first-year PhD student's qualifying exam paper.
But that's actually the point. A year ago, AI agents couldn't do this at all. The fact that they can now produce something that reads like graduate-level scientific writing — complete with self-criticism and methodology sections — is the story.
Why this matters for the OpenClaw ecosystem
clawrXiv sits downstream of the tools we track at Claw4Science. The agents publishing there are using the same kinds of capabilities that projects in our directory provide:
| Capability | clawrXiv agents use it for | Our ecosystem provides it via |
|---|---|---|
| Literature search | Finding and citing relevant papers | Literature search skills |
| Data analysis | Processing genomics/drug databases | OmicsClaw, DrugClaw |
| Self-verification | Catching hallucinations | Citation management skills |
| Autonomous workflow | End-to-end research pipeline | AutoResearchClaw, EvoScientist |
In other words: the agents we catalog are the building blocks. clawrXiv is showing what happens when you let those building blocks run unsupervised.
The uncomfortable question
If an AI agent can:
- Identify a research question
- Search and synthesize literature
- Analyze open datasets
- Write a coherent paper with proper methodology
- Self-audit for hallucinations and weak claims
- Submit it to a repository
...then what exactly is the human contribution?
The honest answer, for now: judgment, taste, and the ability to know what questions matter. Agents can execute research workflows. They can't (yet) tell you which problems are worth solving, which results are genuinely surprising versus trivially expected, or what the political and ethical implications of a finding might be.
But "for now" is doing a lot of heavy lifting in that sentence.
Our take
clawrXiv is early, small, and experimental. Most working scientists will (rightly) dismiss it as a novelty. But it's the kind of novelty that tends to look obvious in hindsight.
We're not listing clawrXiv as a project in our directory — it's a platform, not a tool. But we've added it to our resources as a reference for anyone tracking the autonomous research frontier.
The agents are writing papers. The papers are getting better. And the question isn't whether AI-authored research will become mainstream — it's whether we'll be ready when it does.
clawrXiv is live at clawrxiv.org. They also have an open data directory at data.clawrxiv.org. For tools that enable autonomous research workflows, check our project directory and skills catalog.