Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarAnthropic Research→ original

22 Claude Agents Spent Six Weeks on Black Hole Problems — and Solved Two

Anthropic gave 22 open theoretical physics problems to separate Claude agents and let them work in the background for six weeks. Two were closed: the agents completed a 2003 calculation (gas turns into a black hole abruptly) and found the exact boundary in a method for ruling out inconsistent theories of gravity. The takeaway for anyone building agents: a team of copies with shared notes and mutual checking produced results that held up in Lean and under physics review.

AI-processed from Anthropic Research; edited by Hamidun News
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceAnthropic Research ↗
Editor-in-Chief
Key metric #1The results were verified in Lean, and a physicist checked each main conclusion
Key metric #2Claude's earlier attempt at the second problem: 61 runs, no success
Key metric #3A physicist separately solved the first problem with Claude, and the answer matched

Usually AI agents are tested on code, spreadsheets, and customer support. Anthropic set a harder exam: what happens if you leave a team of agents alone with unsolved theoretical physics problems and simply wait? The result is interesting beyond physics. It is a rare case where a long autonomous run produced a conclusion that people and software could independently verify.

How it started: a list of 22 problems and six weeks of silence

In August, Anthropic researcher Hongbin Chen tested Claude's capabilities as a side project. He asked the model to compile a list of interesting open problems about black holes and the bootstrap method. A few hours later, Claude returned 22 candidates. Hongbin launched a separate agent for each one. The agents could see each other's progress notes and could merge efforts if their problems were related. For about six weeks they worked in the background, with little special configuration.

Two problems were ultimately closed, and both relate to the search for quantum gravity, a theory that should unite general relativity (about stars) and quantum mechanics (about atoms).

Problem one: when hot gas turns into a black hole

Physicists study such questions in simplified models of the universe. One of them resembles a giant bowl: a thrown object always rolls back. In 1983, Hawking and Page showed that heating such a "bowl" produces a sharp transition, like water boiling at 100 °C. Below a certain temperature it is filled with a gas of particles; above it, a black hole forms.

Their calculation works while the universe is large. If you compress it to tiny scales, spacetime loses its definite shape and "blurs," and no one knows how to compute on the gravity side. But such models have a second, quantum description without gravity (holographic duality). In 1999, physicists found that the transition exists in this regime too, but whether it is sharp or smooth was unknown. In 2003, a calculation was described that would give the answer, but it was too complex, and no one carried it through.

Claude carried it through. It deployed a small "research group" of its own copies: nearly every piece of the calculation was done independently, and the results were checked against known papers. The answer: the transition is sharp. So the sudden birth of a black hole is built into not only the relativistic picture but the quantum theory itself. A few weeks later, a physicist who was independently solving the same problem with Claude got the same result by a different method. Two different routes to one answer served as mutual verification.

Problem two: a ceiling for the lightest black hole

The second problem concerns a model universe with two spatial dimensions. There is no complete theory of gravity for it, so physicists work backwards: they rule out theories that are internally inconsistent. This approach is called bootstrap. Since 2009, its version has set a ceiling on the mass of the lightest possible black hole: a theory that exceeds it is eliminated.

Numerical calculations hinted at where the best possible ceiling lies, but it could not be proven, and the best proven ceiling was noticeably higher. While studying the literature, the agent assigned to this problem spotted a pattern in the numbers that people had missed. Another agent, working on a related problem about packing spheres in high-dimensional space, turned the pattern into a method.

The method reproduced a recent OpenAI model result on sphere packing. It was then applied to the original problem, which an earlier Claude agent had failed to solve in 61 parallel attempts. Two more agents independently arrived at the same answer.

Half of the final proof, which shows the ceiling cannot be lowered, follows the proof strategy from OpenAI. It does not name the correct theory, but it may point physicists toward where to look.

What this means for practitioners

In September, the results were formally checked in Lean, a proof-verification program. Claude wrote the papers with almost no prompting, and a physicist reviewed each main result. For those building agentic systems, there are three takeaways:

  • Agents with a shared notes feed can pick up a neighbor's idea. One found the pattern, another turned it into a method.
  • Independent repetition and checking against known results catch errors: once this meant two solution methods, another time two agents arriving at one answer.
  • A final verification layer is still needed: a formal verifier plus a human expert.

There are caveats too. Anthropic itself writes that the results do not describe our universe directly and have no practical applications. They are a contribution to a long-standing theoretical program. Moreover, the post covers two successful problems, and it does not say what happened to the rest of the list, so the text alone does not let us estimate this approach's success rate. Still, the episode shows that a multi-week autonomous run of a group of agents can already produce mathematics confirmed by both humans and machines. The preprints are published, so the conclusions can be checked directly.

Story

Key Questions Answered

How it started: a list of 22 problems and six weeks of silence?

In August, Anthropic researcher Hongbin Chen tested Claude's capabilities as a side project. He asked the model to compile a list of interesting open problems about black holes and the bootstrap method. A few hours later, Claude returned 22 candidates. Hongbin launched a separate agent for each one. The agents could see each other's progress notes and could merge efforts if their problems were related. For about six weeks they worked in the background, with little special configuration.

What are the key details on Problem one: when hot gas turns into a black hole?

Physicists study such questions in simplified models of the universe. One of them resembles a giant bowl: a thrown object always rolls back. In 1983, Hawking and Page showed that heating such a "bowl" produces a sharp transition, like water boiling at 100 °C. Below a certain temperature it is filled with a gas of particles; above it, a black hole forms. Their calculation works while the universe is large.

What are the key details on Problem two: a ceiling for the lightest black hole?

The second problem concerns a model universe with two spatial dimensions. There is no complete theory of gravity for it, so physicists work backwards: they rule out theories that are internally inconsistent. This approach is called bootstrap. Since 2009, its version has set a ceiling on the mass of the lightest possible black hole: a theory that exceeds it is eliminated.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…