Assembled auto-approves low-risk PRs with Codex and Claude Code reviewers, lifting PRs merged to 2.4× its pre-agent baseline
After adopting coding agents, Assembled found that human review had become the bottleneck: median time to first review rose from 3.5 hours in December 2025 to over 16 hours in May 2026, and PRs grew 2.5× larger. CTO John Wang describes an auto-reviewer built into 143.dev, Assembled's internal coding agent platform. Codex and Claude Code each review a PR with the built-in /review command in separate cloud sandboxes, then an orchestrator agent applies an approval policy (size limits, testing evidence, no auth/billing/permissions changes, no new architectural patterns). After a phased rollout, PRs merged reached 2.4× the pre-agent baseline, auto-approvals reached 43% of merges, and refactoring rose 7.5×. Bugs reported fell 27% and reverts held at 0.8%.
The numbers (as reported)
| PRs merged | 2.4× pre-agent baseline |
|---|---|
| Lines of code | 2.8× pre-agent baseline |
| Auto-approvals | 43% of merges, three weeks after full-team launch |
| Large PRs (600+ lines) throughput | 3.5×, p90 wait time steady |
| Refactoring / test work | 7.5× / 3.4× |
| Bugs reported per week | Down 27% |
| Reverts per PR | Steady at 0.8% |
| Median auto-review time | 8 minutes |
| Estimated LLM cost | $3.57 per review attempt (14-day sample) |
| Time to first review before auto-review | 3.5 hours (Dec 2025) → over 16 hours (May 2026) |
Facts
- Company / subject
- Assembled
- Tools
- OpenAI Codex, Claude Code, Devin
- Models
- GPT-5.6-sol, Claude Opus 5
- In-house tools
- 143.dev (Assembled)
- Codebase
- brownfield
- Work type
- Company rollout
- Stack
- Graphite
- Published
- 2026-08-07
- Last checked
- 2026-11-10
What the source claims Practitioner (own blog/paper)
Claimed by the builder, writing about their own work. Not checked independently.
- PRs merged rose to 2.4× the pre-agent baseline 2.4× PRs merged
“PRs merged rose to 2.4× of our pre-agent baseline.”
- Auto-approvals grew to 43% of merges within three weeks of full-team launch 43% of merges
“Auto-approvals grew to 43% of merges over the three weeks after the full-team launch.”
- Large PRs (600+ lines) rose 3.5× in throughput with p90 wait time steady 3.5× large-PR throughput
“Large PRs (600+ lines) increased 3.5× in throughput, with p90 wait time holding steady.”
- Refactoring up 7.5× and test work up 3.4× 7.5× refactoring, 3.4× tests
“Refactoring went up 7.5× and test work went up 3.4×.”
- Bugs reported per week down 27%, reverts steady at 0.8% -27% bugs reported
“Bugs reported per week went down 27% and reverts per PR stayed steady at 0.8%.”
Caveats
Self-reported by Assembled's CTO. The author calls these observational rollout results, not a randomized trial, and says the auto-reviewer doesn't get credit for every point. The window is short (three weeks after full-team launch), and the bug figures are explicitly too early to be conclusive. The pre-agent baseline period is not precisely dated, and no absolute PR counts are given. Category multipliers differ between the intro (bugfixes 3.7×, refactors 7.5×, features 2.5×) and the results list (refactoring 7.5×, test work 3.4×). The $3.57 per attempt is an estimate, not total spend. Codex and Claude Code are the reviewers; Devin is mentioned only for non-engineer PRs, and the agents that wrote most of the PRs are not broken down.
More stories by tool: Claude CodeOpenAI CodexDevin
Related topics: Company rollouts and costBrownfield codebases