Tian Pan: the morning review queue turns overnight agent work into reconstruction, not review
Jul 2 Tian Pan: unattended agents move the bottleneck to 9 a.m.; assign trust tiers at dispatch, demand a provenance receipt with a deferred list, and batch the night's questions into one decision queue.
Source: The Morning Review Queue: Triaging Eight Hours of Unattended Agent Work
Tian Pan starts from what overnight agents actually deliver: not finished pull requests but a queue, with six branches, two failed runs, a dependency bump nobody asked for, and a refactor that is either brilliant or subtly wrong. Citing telemetry from more than 10,000 developers, he notes that high-adoption teams merged 98% more PRs while review time rose 91% and PR size grew 154%. The bottleneck did not vanish when agents took the night shift. It moved to the first ninety minutes of the morning.
His fixes are all decided before the night starts. Trust tiers are assigned at dispatch time and earned per task type, not per agent: fifty clean dependency bumps earn unattended dependency bumps and nothing about migrations, and demotion is automatic when revert rates slip. Each run owes its reviewer a compact provenance receipt with the intent as understood, a decision log, real command output rather than "tests pass," anomalies such as retries and self-reverts, and a deferred list of what the agent chose not to do. That last part turns an absence a diff can never show into an explicit claim a reviewer can check. Unattended agents should defer and continue: pick the conservative reading, finish what can be finished, and route the question to a single morning decision queue instead of hiding ten guesses in ten branches.
Why it matters for code review
Pan's key distinction is that reviewing a colleague's PR is verification of a journey you mostly shared, while reviewing eight hours of agent output is reconstruction of decisions you were not there for. Always-on agents only compound if the morning stays short, and that depends on dispatch-time tiers and receipts the harness enforces, not on a reviewer's stamina. Pair with Pan's mixed PR queue throughput piece, Mouse's overnight house rules, and Rahman et al. on evidence and ownership in AI-authored PR review.