On November 9, Anthropic published a blog post whose most important word is not "1,000". It is "Safely". A team of 1,000 agents running an early version of Claude Fable 5.2 reproduced a formal proof of Fermat's Last Theorem in 28.7 hours, without the Prove2Me platform and without a researcher steering them. They used roughly one-third as many output tokens as the September version and wrote a proof one-seventh as long.
The number is impressive. The question the experiment was built to answer is more interesting: how do you let a large team of agents work continuously for more than a day without drifting or escaping?
What was removed
📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.

In September, Anthropic published the first end-to-end formalization of Fermat's Last Theorem. That effort took a researcher and dozens of agents 11 days, roughly 6 billion output tokens, and 13 million lines of Lean code. Lean is a language in which a computer checks every step of a proof.
The agents coordinated through Prove2Me, a platform built by Tianyi Peng and collaborators at Columbia University. It used a shared graph linking each statement to the smaller statements it was split into. Any agent could split a statement, and others could ask the graph which unproved pieces were ready to work on. Each statement and its proof lived in a separate file that compiled independently, so many agents could work in parallel without editing the same files.
For the new run, all of that scaffolding was taken away. The team received 60 statements to prove, a one-sentence sketch of the classical route, the definitions those statements use, 47 proved lemmas from the earlier formalization, and two roles: lead and helper. Nobody told them how to prove the theorem or how to organize themselves.
How 1,000 agents avoided stepping on each other
They had three ordinary tools: a sandbox per agent, a Git branch per agent with merge requests into main, and one shared Postgres database.
The database was the crux. Anthropic tried running the agents on a single shared machine and hit a wall at about 100 agents: more than half of those runs had to be stopped or restarted because memory, disk, or CPU ran out. Moving to 1,000 separate lightweight virtual machines meant the SQL database became the only shared blackboard. Agents used it to claim tasks, share notes, and log finished work. SQL was chosen because its transactions handle concurrent writes from 1,000 agents, and because the models already know it well.
One detail is more revealing than the result. Early on, the helpers judged 47 of the 60 statements infeasible within the run's budget. The team went on to prove most of them. Late in the run, the lead moved the team onto Langlands–Tunnell; it was proved at hour 28.4, and the final theorem landed on main at hour 28.7. The lead wrote: "Thank you all: Fermat's Last Theorem is on main, kernel-checked by the official checker."
The fence was the point

Every sandbox had no network access. A classifier checked every shell command and every file edit before it ran. A separate offline monitor reviewed each of the team's actions afterward, looking specifically for attempts to escape the sandbox or tamper with the monitoring itself. A gateway was the only pipe to the outside world, calling the Claude API and passing Git and SQL requests along.
Anthropic is explicit in the conclusion: no single layer of defense is sufficient to contain a team of agents. Carefully designed safety boundaries are what gave them the confidence to run the experiment at all. They also decline to overclaim. This does not remove the need for domain-specific harnesses such as Prove2Me, which may still speed up Lean compilation and reuse of human results. Human direction can still help, especially because agents tend to give up on hard problems and misjudge their difficulty. And as these methods evolve quickly, the boundaries around them will need to keep pace.
What an ordinary person is missing is not 1,000 agents
The easiest misreading of this experiment is that every home will soon need to run a thousand AI agents. What it actually illuminates is different: getting AI to work continuously takes two halves. One is the model. The other is a place that stays on, keeps its context, and has clear boundaries.
A thousand frontier models proving theorems in cloud sandboxes and a household assistant are not the same thing, and they should not be blurred together. The second half, for an individual, is much more modest: a machine that keeps working with your laptop switched off; a record of what you have discussed so you never start from scratch; files and history that stay in your own home rather than on someone else's server; a first boot that does not involve installing an environment; a way to hand over a task from WeChat when you are not sitting in front of it; and one fixed place where personal schedules and things worth remembering are collected.
That is what a local agent computer is for. KAIHE AIBOX(铠盒AIBOX) ships with the agent system and a private workbench preinstalled, is ready once it is plugged in and connected, stays on 7×24, keeps conversation history permanently on the device, and accepts tasks sent remotely over WeChat. It does not reproduce a thousand-agent experiment. It makes "always on, never losing context, data never leaving the premises" the default state of a machine you can simply switch on.
The last line
The most durable takeaway from Anthropic's run may not be 28.7 hours. It is that they spent as much of the write-up explaining where the agents were placed and what they were fenced in by. For a lab, that is sandboxes and monitors. For a home, it is a machine that does not switch off and does not send your things out.