Overnight agents need a verification layer, not a longer session
Mar 12 Vincent Tran: builder/judge separation, one browser agent per acceptance criterion, artifact-backed promotion — overnight coding without verification is morning cleanup.
Source: Agents That Run While You Sleep: The Verification Layer Autonomous Coding Needed
Vincent Tran’s Mar 12 note (reposted on DEV Jun 15) reframes the overnight problem: generation throughput is mostly solved; the blocker is verification confidence. Without a verification layer, unattended coding produces morning cleanup. With one, it can produce morning momentum.
The pattern is builder/judge separation. One agent implements against a spec. A separate loop validates behavior against explicit acceptance criteria. Promotion depends on pass/fail evidence, not the builder’s self-assessment. Specs that work for this loop define observable behavior, encode constraints, and include ACs two independent agents would score the same way — “dashboard feels fast” fails; “signed-in user with 200 records opens /dashboard under 2s with no console errors” can be gated.
The practical multiplier is one browser-checking agent per AC, run in parallel: parse testable criteria, plan per AC, launch Playwright-backed workers against local/staging, collect screenshots/traces/logs, then a constrained judge that maps artifacts to AC IDs and emits structured pass/fail. Headless builder invocation (claude -p) only stays trustworthy with budgets, tool boundaries, clean workspaces, and stable output schemas. Pre-flight (boot, env, accounts, routes) kills false negatives before the expensive parallel wave. Failure taxonomy matters operationally: spec ambiguity, environment, execution flakiness, and product failure have different owners — otherwise one team drowns in every overnight incident.
Security and rollout stay boring on purpose: least-privilege tokens, blocked exfiltration, pinned deps, branch protection, human approval for auth/billing/infra; start with one slice and three to five high-quality ACs before widening. Opslane’s verify project is cited as a reference architecture for the same shape.
Why it matters for code review
Continuous verified throughput is this stack — independent judge, artifact-backed ACs, deterministic promotion — not a longer peak Claude Code session. FOMO about idle overnight agents is the wrong scarcity if the morning inventory is unproven work; review bandwidth is better spent on diffs that already failed closed without evidence.