Amazon Shows How AI Agents Can Check Their Own Math Proofs
Amazon Science said in a blog post on Nov. 18 that Amazon Web Services has released public tools that let an AI agent write math proofs while a computer checks them. The post is by Hanno Becker and Larry Paulson.
The aim is to deal with a known weakness. AI agents sometimes make things up. In math, that is a problem, because a proof is either correct or it is not. The authors point to proof assistants, programs that read proofs written in a precise formal language and test them. The check can run while the proof is being written, so a wrong step gets sent back for revision. The agent and the proof assistant talk through the Model Context Protocol, which the post says the leading AI agents support.
The post focuses on a proof assistant called Isabelle. Three tools in Amazon's AutoCorrode suite connect it to an agent. I/Q lets an agent work in a live Isabelle session, and the user can hand over control for a while. I/R lets an agent develop proofs while the user keeps working. Isabelle Assistant, which runs on I/Q, gives the user a chat window to an agent that can view and edit the session.
The authors tried it with Claude Opus 4.6, one of the systems in our list of AI models. Using Isabelle Assistant, it produced proofs that the square root of two is irrational, and that for any N there are N consecutive composite numbers. It also handled the integral of sin x from 0 to pi, which equals two. For the composite-number task, the authors gave Claude a flawed proof. They say Claude noticed it covered only N minus 1 integers and fixed it. For the calculus case, the authors asked only for a simple example, and Claude wrote the code and comments.
The authors describe AI as a partner that assists mathematicians, not one that replaces them. The text of the post available here cuts off partway through the calculus example. It does not say how the tools do on research-level problems, or what they cost to run. Machine-checked math is also being tried elsewhere, as in our report on Turing Agent formal verification.
Sources
Cite this
Amazon Shows How AI Agents Can Check Their Own Math Proofs, New Mind Museum, November 18, 2026. https://newmindmuseum.com/research/2026/11/amazon-ai-math-proof-checking-isabelle/
New Mind Museum. (2026, November 18). Amazon Shows How AI Agents Can Check Their Own Math Proofs. https://newmindmuseum.com/research/2026/11/amazon-ai-math-proof-checking-isabelle/
“Amazon Shows How AI Agents Can Check Their Own Math Proofs.” New Mind Museum, 18 Nov. 2026, https://newmindmuseum.com/research/2026/11/amazon-ai-math-proof-checking-isabelle/.
@misc{new-mind-museum-2026-amazon-shows-how-ai-agents-can-check-the,
author = {{New Mind Museum}},
title = {Amazon Shows How AI Agents Can Check Their Own Math Proofs},
year = {2026},
url = {https://newmindmuseum.com/research/2026/11/amazon-ai-math-proof-checking-isabelle/}
}Free to quote and cite with a link to this page.