OpenScience puts a research agent on the desktop, with every step on the record

The newest entry in the crowded field of AI-for-science tooling is a GitHub repository. OpenScience, published on 31 October under the synthetic-sciences organisation, describes itself as "the open-source AI workbench for scientific research" and ships as a desktop app for macOS, Windows and Linux, a browser workspace and a terminal command.

The pitch is narrow on purpose. According to the project's README, the agent handles "the parts of research that are real work but not the idea": pulling and cleaning data, reproducing a claim, sweeping a parameter, drafting the methods section, checking a reference. The user keeps the idea and the judgment.

What distinguishes it from a chatbot wrapper is the audit trail. The README says a turn reads as "what it thought, what it searched, what it ran, what it wrote, then the answer," and that nothing runs that the user cannot see afterwards. The software bundles shell, Python and R kernels, notebooks, a file system with explicit read and write grants, and remote compute. It also lists hundreds of bundled skills across biology, chemistry, physics, ML and data engineering, plus connectors to ChEMBL, UniProt, PubMed and arXiv.

Model choice is left to the user: bring an API key, sign in to a supported provider, run a local model through Ollama or LM Studio, or pay for a managed option the project calls Ace. The README closes with a caution that reads like a liability note as much as advice: review sources, assumptions, code and outputs before relying on a scientific conclusion.

Automation arrives on both sides of the lab bench

The OpenScience release lands in a week of similar signals. Opentrons, the lab robotics company, announced a $40 million Series E round led by Avego on 29 October, with coverage in citybiz and Genetic Engineering and Biotechnology News framing it as a push to standardise AI-driven science on its robots. The same week, A*STAR set up an AI-powered automated lab for advanced materials development, as reported by Chinese-language outlet Lianhe Zaobao, and XtalPi said it would partner with A*STAR under the MARVEL-AI programme, according to PR Newswire.

Research groups are publishing the evaluation work alongside the products. A paper submitted to arXiv on 30 October by Akshat Jani, Sakhinana Sagar Srinivas and Venkataramana Runkana tests whether agents can write complete executable JSON specifications for geometry creation, meshing, PDE solving and result export. On 120 template-related requests, the full workflow reached 98.3% correct first-turn specification and decision outcomes, against 92.5% for a schema-assisted single shot and 82.5% for model-only, gains of 5.8 and 15.8 percentage points. Among 42 requests that required clarification or rejection, it made no false-ready decisions. Of 12 live user-to-results tests, 9 achieved joint specification, execution, export and numerical-reference success.

A second arXiv paper, submitted the same day by Jae-Hyun Baek, Byung-Sun Won and Jon-Lark Kim, uses LLM-guided program evolution to search for entanglement-assisted quantum error-correcting codes, with Lean checking the algebraic certificate. The authors report the resulting family closed 54 gaps in archived tables and that they released 115 verified code records and ten nonexistence results with checked DRAT proofs. In eight paired generation blocks, validation-selected iterative programs succeeded in 46.1% and 31.6% of held-out executions at unseen lengths, against 19.5% and 3.1% for independently generated programs.

Not every claim of AI assistance in science survives scrutiny. Science News reported on 26 October that Anthropic had described a researcher using Claude to plan work on highly pathogenic avian influenza, including mutations that could help the virus adapt to mammals. Influenza researcher Thomas Friedrich of the University of Wisconsin–Madison called such research "absolutely a normal part of virology research," while biodefence researcher Gregory Koblentz of George Mason University told the outlet there is "no evidence that any of this research was connected to the development of biological weapons." Friedrich also said Anthropic's own account suggests the model "did not meaningfully add scientific capability."

Elsewhere the week's AI news is about cost, not capability. SUSE told its 1,900 employees it will cut about 10 percent of its global workforce, citing "Automatisierung und KI," according to an internal email obtained by The Register and reported by heise online on 2 November. CEO Dirk-Peter van Leeuwen wrote that the decisions are "schmerzhaft für unsere Mitarbeiter, unsere Unterkultur und unsere Open-Source-Community" but necessary for a resilient SUSE. Affected staff were to be informed by Wednesday, with individual conversations starting early next week and a strategy briefing for the rest of the company on 16 November.

For the open-source workbench crowd, the lesson is the same one the README states plainly: the agent shows you what it did so that you can check it.