When agents outpace reading: review evidence, not lines

Chandan KS (Zoroval) starts from a volume problem: coding agents write most of their code, median PRs are ~133 lines, about one in eight exceeds 1,000 — and reading every line or blocking until someone has time both fail. Their third path is review evidence, not lines. The failure that forced it: on July 13 their LLM-as-a-judge missed every known failure in a 1,138-call monitoring run. The model and prompt template were unchanged; the context builder had drifted, and evals only fired on model or prompt changes.

The Intelligence Change Gate is the mechanism. A policy lists protected paths (prompts, context builders, model settings, retrieval). CI fingerprints the exact protected diff. A matching eval report must name that fingerprint, baseline, candidate, and per-stage results; the gate recomputes pass/fail from the numbers. Any later edit to a protected file changes the fingerprint, so old evidence dies automatically. Owner overrides are per-PR, named, expiring, and shown as owner_override, never as a pass. Gate failure output is written as a work order for the next agent turn. Production approval binds to an exact SHA; agents must not own the gate files (CODEOWNERS + required checks). Staging holds 43 eval reports; overrides were used on 9 PRs.

Why it matters for code review

This is how continuous agent throughput stays reviewable without pretending humans will read every line: gates talk to agents, evidence is bound to exact content, and the human reads a one-page packet instead of a thousand-line diff. It productively complicates “always-on factories” too — a one-line context-builder edit can demand a live eval, and the gate only proves what the golden dataset covers. Their open skill packages the setup for Claude Code or Codex: github.com/Zoroval/agent-gates. Compare DispatchSEO’s path allowlist as the real gate and Firmo’s proof-as-merge requirement.