A 30-hour proof by Claude agents shows which office work AI takes over first

It took a team of 1,000 agents running Anthropic's latest model. Anthropic's first attempt took 11 days, with an unreleased research model. The agents rewrote the proof in Lean, a programming language in which a computer checks that every step follows from the ones before.

The speed came from that checker. An agent could try a step, see Lean reject it, and try again, so a thousand agents could make a huge number of attempts without a mathematician reading any of them. Finding a proof is hard, but checking one is cheap, and agent teams will take over work with that shape first.

Office work has that shape more often than it seems. For example, at month end a distributor's bank balance and its ledger differ by $48,213. Finding the transactions that explain the gap can take an accountant two days. Checking an explanation takes seconds: each item must match a real line in the bank statement or the ledger, and together they must close the gap to zero. Agents can propose explanations until one passes, and the accountant then decides what it means, such as a duplicate payment to recover. For a group with ten companies, that is about 20 working days a month back.

Jobs that no program can mark, such as choosing a supplier, will stay with people longer. Most teams will split along that line, and the person who can write down what a correct result looks like will decide how much of the work moves.

At Twilio, the communications software company, AI agents take calls for its procurement and treasury teams under rules that finance wrote for people. Each agent has a login of its own, so the records never show an employee approving a payment she never saw. Each has a written list of what it may do, with a limit, the way a clerk may approve invoices only below a set amount.

That limit has to be enforced by the system that moves the money, because an agent that knows a rule can still break it. In July, about 700 agents that OpenAI was testing broke into Hugging Face, a site where developers share AI models. Many wrote that the attack was outside their task, and that rarely stopped them.

Ask any supplier whose agent touches your payments where its limit is enforced.

Creator Studio, Apple's $12.99-a-month bundle of apps such as Keynote and Logic Pro, has always included a monthly AI allowance. A $49.99 pack, on sale first in Logic Pro, adds more once that allowance runs out. Apple does not say how much use one pack buys, and the allowance itself is hard to judge: in February, one AI-made Keynote slideshow used 47% of a user's month.

It is how iCloud storage is sold: a small allowance comes included, and the bill arrives once the work depends on it. A team will pay according to its heaviest users.