When agents outrun review: accountability, WIP limits, unlike blind spots

László Szabó’s Sep 22 field note is the rare practitioner piece that treats the review bottleneck as accountability and flow, not just “buy another review bot.” Agentic coding lifts the implementation ceiling and leaves review capacity where it was — five thoughtful reviewers still five — so the pipeline becomes idea → very fast development → a large review queue → merge when someone gets to it. In a cybersecurity / critical-infrastructure setting, management pressure to match feature velocity to agent speed collided with engineers asking a non-technical question: who takes responsibility for code no human has ever reviewed?

The diagnosis of broken incentives is sharp. Commit counts and merged-PR counts were already bad KPIs; under agents they become near-free and push cost onto the stage that did not get cheaper. Agent-to-agent approve looks attractive on throughput, but two agents can share blind spots — same misread requirement, same convincing wrong explanation — and neither carries consequence. The redesign sketch is familiar and useful: automated verification before humans, an independent review agent, risk classification that sets review depth, human judgment where it matters, merge queue against the real base. Push agents toward smaller independently reviewable PRs, and land pre-processed artifacts (what/why/tests/assumptions/hot files) rather than raw 1,500-line dumps. Unreviewed PRs are inventory: example WIP limit of three pending PRs per engineer or agent; past the ceiling, agents switch to answering review feedback, tests, docs, design — not opening more queue.

Honest product status for vincent (his local control plane): deterministic check gates the agent cannot talk past, manual publish gates, pre-assembled gate instructions, PR+CI in one screen with merge unreachable via MCP. Still open: blast-radius risk classification, a first-class independent reviewer with unlike model/context/prompt, and WIP limits on pending PRs (today only on running tasks).

Why it matters for code review

Continuous harnessed throughput is not more overnight commits — it is a pipeline that caps unfinished review inventory and refuses to treat same-model rubber stamps as accountability. Measure cleared review and wait time, not opened PRs; require independent review that does not share the implementer’s blind spots; keep humans on architecture and risk while scripts eat the mechanical queue. That is how always-on factories compound past burst mastery without turning FOMO into a merge KPI nobody can defend after an incident.