Skip to main content
Back to timeline
Amazon Science homepageSource publication:

AWS plugs AI agents into the Isabelle proof assistant: proofs are checked as they are written, not trusted on faith

Synopsis

AWS has released three tools in its AutoCorrode verification suite, I/Q, I/R and Isabelle Assistant, that let an AI agent talk directly to the Isabelle proof assistant through the Model Context Protocol. The proof assistant checks each formal proof step by step as the agent writes it, so generation and checking alternate instead of relying on trust in the model's output. In demonstrations, Claude Opus 4.6 with Isabelle Assistant formalized the irrationality of the square root of two, blocks of consecutive composite numbers, and the definite integral of sin x over [0, π], and it corrected a supplied proof that covered only N-1 consecutive integers.

AI-generated editorial illustration: Mathematical proof, powered by AI. And no hallucinations!

Interpretation

An AI agent can now operate a live Isabelle proof-assistant session directly, turning a mathematical task stated in ordinary terms into a formal proof a machine can check. Proof assistants were built mainly for human users to write and check proofs, and agents lacked a simple, direct way in; most assistants expose sophisticated human-facing interfaces, such as the Lean extension for VS Code. Among three examples completed automatically with Isabelle Assistant, the proof that the square root of two is irrational was called surprisingly idiomatic though more detailed than strictly necessary, and the integral proof was a bit longer than necessary but error free.

Correctness is decided by the proof assistant checking each step, not by a person reading the agent's output and deciding whether to believe it. Agents are notorious for making things up, so their solutions normally have to be checked carefully by hand. Interleaving generation with checking keeps a proof on a checkable track as it is produced; the author credits this approach with making the integral proof error free.

I/Q lets an agent temporarily take control of a proof-assistant session, I/R lets an agent develop formal material in a background subprocess without touching the user's document, and Isabelle Assistant wraps I/Q as a chat window inside the user's session. A command-line interface is enough to connect an agent to an assistant such as HOL Light, but not to a system like Isabelle with a sophisticated user interface. All three connect agent and tool through the Model Context Protocol and are publicly available in the AutoCorrode verification suite.

Perspective

The demonstrated scope is Isabelle, with three examples, the irrationality of the square root of two, blocks of N consecutive composite numbers, and the definite integral of sin x over [0, π], all completed automatically by Claude Opus 4.6 with Isabelle Assistant; the author explicitly calls these examples easy and says harder problems need I/Q and I/R.

The author says the hardest mathematical problems still require human thought, and how this new kind of teamwork between agent and human works best remains to be figured out; generated proofs often contain complications and detours a human would avoid, which a partially automated polishing phase can remove.