When an AI agent writes a smart contract, reviewing every line of generated code can become a bottleneck. An article published on November 17, 2026, proposes a different starting point: write the properties a change must satisfy in English, have a person review those properties, then use formal verification tools to check whether the code meets them.
The article describes a demonstration by Certora CEO Mooly Sagiv involving an ERC-4626 USDC vault. An agent inferred properties for the vault and generated Solidity code and CVL verification rules. According to the article, the prover found a dilution issue of about 5 wei caused by rounding in the high-water mark; the issue was then fixed.
That example shows why a passing code review is not the whole task. Formal verification checks code against a specification—the stated properties it is meant to satisfy—but the article identifies writing and assessing that specification as the hardest part. Its proposed workflow puts human review at that point, before relying on the prover’s result.
