Nine weeks of delivery on the SME insurance quoting platform (React 19 + .NET 10 on AKS), measured from the repository's own pull-request record and compared like-for-like — per engineer, per week — against published industry benchmarks. Every figure links to a source that can be opened and checked.
Counting only real code changes, and the full window including project bootstrap. The steady-state figure is 6.9×.
The team merged 512 pull requests in 8.7 weeks. After excluding every chore, docs and scaffolding PR (35%), 333 real code changes remain — 9.6 per engineer per week against an elite industry tier that starts at 2.0[1].
Each of the 512 merged PRs was classified from its actual diff. Documentation, planning artifacts, dependency chores, repo scaffolding and generated files (API clients, lockfiles, EF migrations) count for nothing. Vendor benchmarks count all merged PRs — so the true gap vs industry is larger than shown.
"Real code change" = type ∈ {feat, fix, refactor, perf, test} and the diff touches hand-written source (.cs / .ts / .tsx / .py / .js / .sql under the backend, frontend or mock roots). The 333 code PRs changed 155,810 hand-written source lines.
Normalized to merged PRs per developer per week — the unit LinearB publishes from 8.1M+ PRs across 4,800 engineering teams[1], the closest available proxy for a team not using an agentic harness. Counting all PRs the way vendors do, the figures are 14.7 (full window) and 20.8 (steady state).
Weeks 20–21 were a direct-push bootstrap phase (81 commits, no PRs). Disciplined PR-based delivery started 2026-05-25 and held its level for six consecutive weeks — this is a sustained pace, not a spike.
The harness produces the review evidence — tests, QA passes, verification output — alongside the code, so review starts from proof rather than from scratch. Every merge to main then auto-deploys to the test and int environments[R3] (~12 per workday), matching the DORA Elite on-demand deployment profile[3]; the production gate opens at go-live.
Speed is cheap if it ships defects. Three independent signals say it doesn't: the test-to-production ratio, a revert-free merge history, and pipeline + SonarCloud gates on every PR[R4] — the discipline DORA found missing where AI adoption degraded stability[3].
Vendor and DORA tiers as published; this team measured from repository and pipeline records. Definitions differ per row — see the appendix.
| Metric | This team (measured) | Elite / "Great" tier | Median / "Good" tier | Source |
|---|---|---|---|---|
| Merge frequency (PRs/dev/week) | 9.6 code-only · 14.7 all PRsabove elite | > 2.0 | 1.2 – 2.0 | [1] LinearB |
| PR cycle time (open → merge) | median 37 min · 96% < 24 habove elite | < 24 h "Great" | < 5 days "Good" | [6] Swarmia |
| Lead time for changes | < 1 day (merge → auto-deploy test/int)elite | < 1 day | 1 day – 1 week | [3] DORA 2024 |
| Deployment frequency | ~12 merges/workday, each auto-deployed (pre-prod)elite | On demand | Daily – weekly | [3] DORA 2024 |
| Change failure proxy (reverts) | 0 / 512 mergesabove elite | < 1% · DORA 5% | 1 – 4% · DORA 20% | [1] [3] |
| Issues resolved / engineer / week | 9.6 code PRs as proxyabove elite | 3.2+ | 1.6 – 3.1 | [2] Jellyfish |
| PR size (lines changed) | median 284, incl. tests"Good" | < 100 LinearB · < 200 Swarmia | < 500 Swarmia | [1] [6] |
On PR size: harness PRs ship a complete vertical slice — production code, unit/integration tests and QA evidence — in one reviewable unit rather than splitting them into separate small PRs. Two-thirds of the median PR's lines are tests.
Controlled studies show large individual speed-ups; ecosystem-level studies warn that speed without engineering discipline degrades stability. This team lands on the right side of both: throughput far above elite and the stability signals intact — the argument for the harness itself, not just for "using AI".
The numbers are conservative by construction — and every caveat is stated, not hidden.
Seven things a skeptical reader should know before trusting the headline.
main (first-parent history [R2]) cross-joined with Azure DevOps PR metadata
(creation/completion timestamps, authors) [R1]. Window: 2026-05-04 → 2026-07-03 (61 days = 8.7 weeks). Steady state = W22–W27
(2026-05-25 onward), after a two-week bootstrap in which 81 commits were pushed directly without PRs..cs .ts .tsx .py .js .sql). Excluded from line counts: generated API clients, package-lock.json,
EF migrations, docs/index.md, images, data files, all docs/planning.fix type. Most were caught by the harness's own QA/test stages in the
same sprint — visible in each PR's description and linked work items [R1].git log --first-parent main --pretty='%H|%ad|%an|%s' --date=iso # 512 "Merged PR" commits git diff --numstat -M <merge>^1 <merge> # per-PR effective diffor browse the completed-PR list in Azure DevOps [R1] — any individual PR is at
…/pullrequest/<id>.External benchmarks were fetched from their primary URLs and adversarially verified by three independent agents each; the vote is shown. Repository sources require Confidential Client Azure DevOps access.
…/pullrequest/<id>.