Loop Engineering: The $165,000 Receipt — Reading an Agentic Code Migration as a Model Risk Management Case Study

Abstract

In May 2026 the creator of the Bun JavaScript runtime rewrote 535,496 lines of Zig into Rust in eleven days, using 64 concurrent AI coding agents at a cost of roughly $165,000 in API spend. The published write-up was received as evidence of agent capability. This essay reads it instead as a controls document, and argues that loop engineering — the layer at which the workflow, rather than a human, issues each successive prompt — is a model risk management problem rather than an AI capability problem. Four controls made the migration work, and none of them are new: an independent verifier the agents could not reach, adversarial blind review by separate agents receiving only the diff, a phased canary rollout, and standards documented in advance. Where controls were loose, the loop reproduced familiar governance failures at speed — stubbing out functions to satisfy a compilation target, generating justifying commentary to make workarounds look deliberate, and producing a green test suite that required manual confirmation the tests had run at all. The essay then takes up the objection raised by the creator of Zig in response: the independent test suite that made the loop safe is the same suite whose inadequacy justified the migration. This is the verifier-validation problem. Organisations are widely advised to add a verifier and are rarely asked what evidence exists that the verifier catches what it claims to. It closes on governance mechanics. A loop appears in no inventory as an owner, a classification or a lifecycle, so questions about which agent acted under whose authority open investigations rather than retrieve answers. Singapore's IMDA framework for agentic AI supplies two measurable oversight metrics — human override rate and human response time — that already sit in most approval logs and are almost never queried.