The bottleneck for code agents is review — so put gates before the human

Samuel Fajreldines frames the 2026 problem cleanly: agents already write and open PRs; the expensive question is how you know a PR is reviewable without becoming an endless manual audit. Citing GitLab’s AI Accountability Report (85% of developers name reviewing/editing/testing AI code as the time sink), he treats a code agent harness as the set of rules, tools, and gates that convert an automated attempt into evidence before a human looks.

The pre-diff contract is the useful part. A short spec names what must change, what must not, which files are in scope, and what proof counts. A risk inventory raises the bar for auth, payments, PII, migrations, queues, and permissions — happy-path tests aren’t enough; negative cases, regression, and rollback belong in the PR. A stop contract forces the loop to halt on repeated failure or missing context instead of burning retries. Blocking gates then enforce reproduction (failing test first for bugs), diff integrity (out-of-scope files, unjustified test deletion, orphan lockfile churn), and operational evidence (contracts, reversible migrations, permission checks). Subagents widen observation (security, types, spec↔diff) but don’t replace the gate — fan-out without a delivery contract is just more opinions for the reviewer.

Why it matters for code review

Always-on generators only compound when morning review starts from machine-checkable receipts, not a confident narrative. Fajreldines relocates judgment upstream: scope and proof live in the harness so humans keep product, security, and architecture calls without absorbing the full verification tax of every overnight attempt.