Six months of agent-driven delivery: 10x merged PRs, and the gains came from handoffs and scheduled checks, not prompts
Oct 3 Vinod Kurien Alex, an engineering lead at a B2B SaaS company, went from 11.6 to 113.6 merged PRs a month with 3 to 4 reverts in a year. He credits a layer of handoffs, a router, scheduled routines and review agents, and is candid that the fair range is 4x to 10x.
Source: Moving an engineering team to agent-driven delivery: six months of real numbers
Vinod Kurien Alex leads engineering at a multi-tenant B2B SaaS company and moved his own work, and much of his team's tooling, onto Claude Code between April and October 2026. Counting only his own commits and PRs, before versus after a May 19 split:
- Merged PRs: 11.6 → 113.6 a month.
- Commits reaching a main branch: 39 → 167 a month.
- Test cases added: 473 → 1,347 a month.
- Median PR cycle time: 2.0 h → 0.7 h (p75 8.7 h → 3.4 h).
- Reverts: 3 to 4 in twelve months.
He is unusually careful about what those numbers don't show. His role grew, the repos he touched doubled, and agent PRs are smaller (lines rose about 4x while PRs rose about 10x), so he calls 4x to 10x the range he can defend. Faster cycle times "partly reflect agents removing waits between people. They don't show that reviews got faster."
His claim is that the gain came from what he built around the model, not from better prompts:
- Handoffs. Every long-running work stream has a charter, and every session ends with a written handoff so the next one can start cold. About 190 handoffs across 14 streams.
- A router session that never edits code: it classifies a request, writes a self-contained brief and dispatches it.
- Scheduled routines: a weekday-morning brief on open work, the PR queue and deploy drift; a daily test-balance audit; a weekly build-and-deploy health check.
- Review agents and a read-only architecture reviewer that checks repos against auth, tenancy and permission rules, plus a lint hook that enforces test conventions on every agent edit.
The failures are the useful half. A scheduled brief posted his private work items to a team channel four times, and a run that had already started kept the old instructions after he fixed the prompt. The router he built to prevent mega-sessions became one: 19 of the last 30 sessions, 30% of tokens, one conversation past 800 turns, and work items opened twice as fast as they were closed. And an agent invented a ticket ID that ended up in the changelogs of two published package versions.
Why it matters for code review
This is a rare month-by-month account where the throughput comes from a standing system (handoffs, scheduled checks, reviewers that run every time) rather than from long, intense sessions, and where reverts stayed rare as volume rose roughly tenfold. It also shows where that system needs its own review. A scheduled agent with write access is a publishing system, an orchestrator drifts the way a team does, and any fact that leaves the repo (an ID, a number, a root cause) needs a check against its source before a reviewer can trust the diff. Pair with the Claude Code team's risk-tiered review and falling reverts, Addy Osmani on state, sessions and handoffs, and Kassebaum's agents setting a risk tier from the diff.