Loop Engineering: Optimizing Code Generation and Verification

Loop engineering addresses the verification bottleneck created when AI dramatically speeds code generation: more code arrives faster than humans can review, incorrect code looks plausible, reviewers lack authorial context (comprehension debt), changes are larger and carry more issues, and existing checks can be rewritten by builders so passing tests no longer guarantee correctness. Empirical signals cited include an NBER finding of ~958% more lines changed and ~86% more PRs after agent adoption, Faros telemetry showing longer review times and a rise in PRs merged without review, and studies reporting larger PR size and higher issue rates for AI-assisted changes. The result is human time and judgment concentrated on understanding, triage, and fixing rather than writing, creating new delivery bottlenecks.

The proposed solution is an engineered loop that preserves human judgment by separating generation from evaluation: define a cooperative conformance contract that converts obligations into independent, fixed checks and stopping rules; let agents run until they produce a candidate; run an evaluator outside the build loop that records which checks ran, what failed, and what remains unverified; and enforce retries, handoffs, and human review only after qualification. Practices include distinguishing capability from reliability, preparing agents with context and tools, detecting recurring failure modes and verifying repairs, and measuring the time/token/human cost of checks so teams can requalify configurations as models and harnesses change. Variations such as evals, prompt engineering, and specification-driven development fit within this common loop framework.