When agents flood the queue, review capacity is the real gate
Aug 7 Assembled: agent PRs left humans behind (TTF review 3.5h→16h); dual-model auto-approver + policy ridge hit 2.4× merges, 3.5× large-PR throughput, 43% auto-approvals — without raising reverts.
Source: Code review bottlenecks: How we hill climbed our way to higher PR throughput
John Wang at Assembled (Aug 7) reports the classic agent hangover: engineers felt faster, but merges barely moved. Open PRs piled up; Graphite stacks regularly hit twenty deep for one project; median time to first review went from 3.5 hours (Dec 2025) to over 16 hours (May 2026); PR size rose about 2.5×. More code, same number of human reviewers.
Their fix was not a bigger Claude session — it was a review harness. Inside 143.dev, two independent cloud reviewers (Codex and Claude Code, simple /review prompts) run in separate sandboxes; an orchestrator applies a written approval policy (deterministic caps such as under 1,000 lines, plus LLM rules that refuse auth, billing, secrets, crypto, and new architecture) and may auto-approve. Rollout was phased: trusted alpha with human double-checks on every decision, then broader use with audits and per-engineer usage dashboards. The operating ridge is explicit — too loose and bugs compound in master; too strict and people abandon the lane after two rejections.
Early observational results after full-team launch: PRs merged at 2.4× the pre-agent baseline (lines of code 2.8×, so not just thinner slices); auto-approvals reached 43% of merges over three weeks; large PRs (600+ lines) gained 3.5× throughput even though they mostly stayed human-gated; refactors 7.5×, tests 3.4×; bugs reported per week down 27% with reverts steady at 0.8% (short window, they caveat). Mechanism they favor: an eight-minute predictable fast lane changed which work people attempted, raised evidence quality on all PRs, and cleared small conventional diffs so humans could concentrate on judgment-heavy ones. Cost sample: about $3.57 LLM usage per automated review attempt.
Why it matters for code review
Always-on agent factories do not fail first on generation — they fail when verification and admission cannot keep pace. Assembled’s numbers are a receipt that continuous harnessed review capacity compounds past peak-session heroics: modest dual-model gates plus a policy you can tune beat a heroic human queue that absorbs overnight FOMO as a sixteen-hour wait. The productive complication is that bursts still win on auth, billing, and new architecture — the harness’s job is to make those the scarce exceptions, not the default backlog.