OpenScience ships an open research agent as lab automation money keeps flowing

The repository describes OpenScience as "a research agent with a workbench around it." You give it a goal in plain language, according to the project's README, and it plans, gathers evidence, runs code and experiments, and returns results the user can check.

It runs as a desktop app, a browser workspace or a terminal command, on the user's machine, against the user's files. The wording on the project page is unusually blunt about what it is not. "It is built for 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 README says. "You keep the idea and the judgment." The design choices read like an answer to a specific worry: that agents doing science are hard to audit.

Every turn shows what the agent thought, what it searched, what it ran and what it wrote.

Shell, Python and R kernels are available, plus a file system with explicit read and write grants and remote compute for jobs a laptop cannot carry. Bundled skills cover biology, chemistry, physics, ML and data engineering, with connectors to ChEMBL, UniProt, PubMed and arXiv. Users bring their own API keys, run a local model, or pay for a managed option called Ace.

The funding line keeps getting longer

OpenScience is not an isolated release. On 29 October, Opentrons announced a $40 million Series E round led by Avego to expand AI-driven laboratory automation, according to Business Wire. Genetic Engineering and Biotechnology News covered the same round on 30 October as money for robotics, software and data infrastructure. Robotics & Automation News followed on 30 October, and hi-Tech.ua reported on 2 November that the company wants GitHub-style collaboration for lab robots.

Capital is also moving into the security and operations side of enterprise IT.

Tech.eu reported on 3 November that Milan-based CyberKord raised €2.7 million in pre-seed funding, led by 360 Capital with Plus Partners, Addendum Capital and Vento participating. The company, founded in 2026 by Alessandro Maggio, Dragos Tanasa and Cosimo Michelagnoli, uses AI and a digital twin of a customer's infrastructure to analyse issues, simulate fixes and apply them. It is not a lab company. It is the same pitch, though: take the manual work out of the loop and let software do the first pass.

Meanwhile the biggest signal this week came from a software vendor cutting jobs and naming automation as the reason. SUSE chief executive Dirk-Peter van Leeuwen told staff in an internal email that the company would cut about 10 percent of its 1,900-strong global workforce, as first reported by The Register on Monday and covered by heise online on 2 November. "Automation and AI are fundamentally changing how software companies work," he wrote, according to heise. Affected employees were to be told by Wednesday, with personal conversations starting early next week.

What the research papers actually show

The technical literature is more cautious than the funding headlines.

On 30 October, researchers led by Jae-Hyun Baek posted a paper on arXiv describing an LLM-guided pipeline for discovering families of entanglement-assisted quantum error-correcting codes. The pipeline combined program evolution with exact evaluation, formal proof checking in Lean and certified nonexistence checks. It closed 54 gaps in archived tables and released 115 verified code records. The authors also ran a controlled comparison: validation-selected iterative programs succeeded in 46.1 percent and 31.6 percent of held-out executions at unseen code lengths, against 19.5 percent and 3.1 percent for independently generated programs.

A second arXiv paper, submitted the same day by Akshat Jani, Sakhinana Sagar Srinivas and Venkataramana Runkana, asks whether agents can specify engineering simulations reliably. On 120 template-related requests, their full workflow reached 98.3 percent correct first-turn specification and decision outcomes, against 92.5 percent for a schema-assisted single shot and 82.5 percent for a model-only approach. The workflow made no false-ready decisions across 42 requests that required clarification or rejection. On 12 live user-to-results tests, 9 achieved joint specification, execution, export and numerical-reference success, or 75 percent.

"Our goal was to examine how different quantum learning approaches perform when the amount of training data is deliberately constrained," said Arslan Munir of Florida Atlantic University, according to HPCwire.

HPCwire reported on 2 November that Munir's team tested quantum machine learning on lung cancer prediction from limited data. D-Wave's annealing machines, used in a quantum support vector machine, reached 94.31 percent accuracy in simulation and 91.90 percent on hardware. The dataset started at 309 records and was balanced to 540 samples with synthetic examples, which the researchers say did not represent newly recruited patients.

Hardware is moving in parallel. Semiengineering reported on 2 November that IBM researchers demonstrated four long-distance couplers between two 120-qubit Nighthawk R1 superconducting processors, in a paper on arXiv. IBM's own work on magic state cultivation, posted to arXiv, points the same way.

Where the scepticism belongs

The case that best captures the gap between demo and deployment is Claude and the bird flu researcher.

Science News reported on 26 October that Anthropic said a researcher used Claude to plan studies and analyse data involving highly pathogenic avian influenza, including work to identify mutations that could help the virus adapt to mammals. Anthropic banned the account in May 2026 for evading access restrictions. Gregory Koblentz of George Mason University told Science News there is "no evidence that any of this research was connected to the development of biological weapons," and Thomas Friedrich of the University of Wisconsin–Madison said the company's account suggests the model "did not meaningfully add scientific capability."

That is the test every one of these tools now faces. Automation can compress the boring parts of research, and the open-source workbench makes its steps inspectable, which is more than most. But a faster loop is not the same as a better question. SUSE's email is the reminder: the same argument that funds the tools is already costing people their jobs.