Black Box’s PR gates — five reviewers before a human merge
Field note from Black Box via Codacy — ~8 coding agents in the loop, every PR through five independent reviewers, plus Verity’s markdown knowledge graph so later turns inherit what got rejected.
Source: Agentic SDLC Loop Engineering: How Black Box Runs PR Review Gates at Scale (2026) — Codacy
Erik Jost at Black Box is running roughly eight agents in the development loop. Every pull request still passes five independent reviewers before he merges by hand: cross-model Claude and Codex for logic and intent the writing model missed, plus Copilot, Codacy, and CodeRabbit, with GitHub’s “all conversations must be resolved” sitting on top as a hard gate.
The useful design choice is independence. Same-model review inherits the writer’s blind spots; cross-model and multi-tool review is how you buy coverage when PR volume is agent-driven. Humans stay on the merge button — the scarce resource is still judgment — but the stack in front of that button is engineered as parallel filters, not one vibes pass.
Why it matters for code review
Codacy’s Verity piece of the story is the compounding layer: an in-loop adversarial review that pairs deterministic checks with an independent model, and writes decisions into a git-tracked markdown knowledge graph so later turns inherit what already got rejected. When agents raise the throughput, review stops being a single tool and becomes a gate architecture — independent filters, memory of past decisions, human at the end.