How One Developer Built a Self-Checking AI System That Refuses to Trust Itself

A developer has built a nightly automated architecture pairing two AI agents — LOOM, an effect-typed language that makes code permissions provable, and ARGUS, a system that treats its own conclusions as suspect until verified. The two agents are deliberately separated so each acts as an independent adversary to the other's work, preventing the blind spots an author inevitably has when reviewing their own code. All information passed between agents must be expressed as hard measurements — hashes, test counts, return codes — rather than narrative claims, making the audit trail machine-checkable. A single human operator retains ultimate authority, including the sole right to declare the current date, closing a class of drift where automated systems quietly redefine their own timelines. The design was validated in practice when one agent caught a real bug in the developer's own backup code that even a misleading source comment, written by the developer themselves, had obscured.