AI Agents · Governance · R&D
From AI Agents to AI Organisations: Who Is Responsible When 60 Agents Work as One?
Anthropic’s experiment with 60 autonomous AI agents offers a glimpse of a new R&D model: one human supervising a coordinated system of agents that can search, reason, challenge results and formalise its work. The technological achievement is striking. For lawyers and organisations deploying agentic AI, however, the more important question is what happens to responsibility, oversight, traceability and accountability when AI begins to operate as a team.
Complex law. Clear action.
Reviewed by Oleksandr Sobovyi, Founder & CEO of CORVUS AI — editorial responsibility statement below.
One person and 60 autonomous agents. A new type of R&D that Claude just demonstrated — and a few legal questions that follow from it
On August 10, Anthropic published a rather understated post. No fanfare. No claim that “we solved the Riemann hypothesis.” Just a straightforward statement: an unreleased research version of Claude, given the task of taking a serious stab at the Riemann hypothesis, did not prove the hypothesis itself. But along the way it produced a new result on a related problem.
The lower bound on the proportion of zeros of the zeta function that lie on the critical line rose from 41.6% to 67.2%.
The previous record had held for years. Progress had been incremental, measured in tenths of a percentage point. Here the jump was more than 25 percentage points in one step. The result was formalized in Lean, reviewed by Anthropic’s internal mathematicians, and examined by two external specialists — Brian Conrey and Dan Goldston.
From a scientific standpoint the story is clear. From a legal standpoint it gets more interesting.
Authorship. In the paper itself the author is listed as Claude. This is not a joke or a marketing flourish. The model proposed writing up the result as a paper and coordinated the Lean formalization. Under classic copyright doctrine (both the U.S. Copyright Office position and most European approaches) an AI is not an author. Rights vest in a human or a company. This raises the familiar work-made-for-hire question: an Anthropic employee set the task, the company provided the model and the infrastructure. The result almost certainly belongs to Anthropic. But if the same work had been done outside the company — say by a university researcher using a public model — the picture could look quite different. Courts have barely begun to address cases of this kind involving serious mathematical results.
Intellectual property in the result itself. A mathematical statement and its proof are generally not patentable subject matter. But a formalized Lean proof, the repository, and the method of organizing a multi-agent search sit closer to protectable artifacts. Who has the right to use, modify, or commercialize that formalization? Anthropic released the materials. That is openness. Legally, however, it is not the same thing as the public domain. License terms, conditions of use, and the scope for downstream development all remain in a gray zone for now.
Verification risk. Two internal mathematicians and two external experts reviewed the work. The Lean check passed. That is strong. Yet from a legal perspective a residual question remains: what happens if an error is found a year from now? Who bears the reputational and, potentially, contractual risk — the company, the employees, the external reviewers? In academia such risks are distributed informally. In a corporate setting, especially when the result is used to position model capabilities, the issue moves closer to questions of disclosure and representations.
Process. A non-mathematician set the task. Roughly 60 sub-agents then worked with a high degree of autonomy: generating ideas, discarding dead ends, checking one another, writing code, searching the literature. This is no longer “a tool in the hands of a researcher.” It is the delegation of a substantial part of the research process to a system. From a corporate governance perspective the question arises: at what level and under what controls may a company allow such systems to produce results that are then published under its name? Especially when the work involves unreleased models whose capabilities have not yet been fully fixed in public descriptions.
One further layer. Anthropic states explicitly that this approach is unlikely to lead to a full proof of the Riemann hypothesis. That is an honest limitation. But the mere fact that the company is publishing a result obtained this way creates a precedent. Other labs and corporations will look not only at the figure 67.2%, but also at the legal scaffolding: how authorship is framed, how rights are allocated, how review is organized, and how limitations are described.
For now this is a single case. Yet the pattern — one operator plus a swarm of research agents — already works. And the law is still catching up more slowly than the mathematics.
What matters. What’s next.
Disclaimer
This article has been prepared by CORVUS AI for general informational and educational purposes only. It is intended to make complex legal and regulatory developments easier to understand.
It does not constitute legal advice and does not create a professional adviser–client relationship. The information should not be relied upon as a substitute for advice based on the specific facts, circumstances and applicable law relevant to your organisation or project.
The article reflects our understanding of the law and regulatory framework as of the date of publication. Legislation, case law, regulatory guidance and administrative practice may subsequently change. While reasonable care has been taken in preparing this article, CORVUS AI does not warrant that the information is complete or remains current after the date of publication. We do not undertake to update this content.
To the fullest extent permitted by applicable law, CORVUS AI excludes liability for loss arising from reliance on this article. Nothing in this article constitutes an offer or solicitation to provide regulated legal services in any jurisdiction where doing so would be unlawful.
AI-assisted preparation: This article was prepared with the assistance of AI tools. Its legal analysis, conclusions and final text were subject to human review and editorial control and were reviewed and approved prior to publication by Oleksandr Sobovyi, Founder & CEO of CORVUS AI. CORVUS AI retains editorial responsibility for the published content.
For advice tailored to your organisation, project or specific circumstances, please contact CORVUS AI.
