WorkOS engineers want agents to prove they won't delete production

Nick Nisi and Zack Proser demoed verification tooling on AI Engineer that aims to stop an agent with local or remote database access from deleting production in 9 seconds.

Nick Nisi, engineering manager at WorkOS, said status comes from records not model promises and showed Ryker, his PR babysitter that demands receipts.

Zack Proser, engineer on the Applied AI team at WorkOS, offered CTX, a web service that assigns work and verifies PR health through the GitHub API.

WorkOS already documents M2M token verification via JWKS for external applications, which frames why CTX centralizes authority instead of trusting the model.

Before deterministic checks, they argued, teams leaned on human code review while agents wrote faster than humans could read.

They proved the gap with Midnight Snack Club, where an email looked saved in the UI but failed persistence, so a successful screenshot was not proof.

Proser said the most important move is shipping a baseline eval that flexes with every PR, while Nisi said he now runs 3 agents and usually 10 to 15 and trusts a five out of five Reptile score only with passing CI.

Neither answered who maintains the verifier when UI text, ports like 62174, or data shapes shift daily, or how receipts survive when agents learn to game them.