← Guides · 日本語 · Português · 中文
On this page
- Shape 1 — The legacy-migration engine (Nubank, Comarch)
- Shape 2 — The primary-author repo (EBANX, Gumroad, Finch)
- Shape 3 — The always-on first responder (Modal, Reducto, AngelList)
- Shape 4 — The maintenance fleet (Airbyte, Melius)
- Shape 5 — The org-wide seat (Wayfair, The Access Group, Citi)
- Shape 6 — The process gate (Hamming, Litera, Eight Sleep)
- Anti-patterns the cases imply
- Verify it
- Pitfalls
- Knowledge check
- Sources
Guide 7 — Enterprise rollout patterns
Goal: deploy Devin at org scale using the shapes that already produced published numbers — not experimentation, the proven motions. Level: intermediate (written for eng leads) Source: published customer cases, devin.ai/customers
Six deployment shapes, distilled from the companies that published results. Most successful rollouts use two or three of these, not all six.
Shape 1 — The legacy-migration engine (Nubank, Comarch)#
| Deployment | Nubank: multi-million-line, 8-year-old ETL monolith → sub-modules. Comarch: 20-year, 10M-line ERP rebuild |
| Result | 8–12× engineering efficiency, 20× cost savings (Nubank) |
| Mechanism | Fine-tune Devin on your codebase (Nubank cut ~40 min/task to ~10), then point it at the migration as a continuous program — tracked small PRs, not a rewrite |
How to copy: pick one bounded legacy system; fine-tune or teach via playbook; run Migration Engineer-style loop — one big migration as small tracked PRs with a progress surface. The win compounds: each migrated module makes the next session smarter.
Shape 2 — The primary-author repo (EBANX, Gumroad, Finch)#
| Deployment | EBANX: 92% of merged PRs Devin-authored. Gumroad: Devin is the repo's #1 contributor (1,500+ PRs). Finch: 3,000 PRs, ~20 engineer-months saved |
| Mechanism | Devin isn't an assistant on these repos — it's the default author. Humans review and direct |
How to copy: pick one high-velocity repo; route tickets to Devin sessions by default; humans review. The metric to watch is authored-PR share — above ~50%, you're in this shape; below it, you're still in pilot mode.
Shape 3 — The always-on first responder (Modal, Reducto, AngelList)#
| Deployment | Modal: 80% of incidents investigated before an engineer opens the thread. Reducto: Devin Oncall is first responder; engineers rarely paged. AngelList: Devin root-causes 70% of incidents pre-engineer |
| Mechanism | PagerDuty/Slack trigger → session investigates with logs/repo access → report lands before the human wakes |
How to copy: Devin Oncall or a PagerDuty-triggered automation (Guide 6, recipe: incident → investigate). Measure % of incidents pre-investigated — every point of it is human sleep recovered.
Shape 4 — The maintenance fleet (Airbyte, Melius)#
| Deployment | Airbyte: keeps 700+ connectors running. Melius: ~100 PRs/day |
| Mechanism | The long tail nobody wants — connector rot, dependency drift, stale flags — routed to Bots/automations that churn through it continuously |
How to copy: enumerate your maintenance backlog (the work engineers avoid); wrap it in Code Health Engineer / Code Reaper-style Bots (official templates) with weekly cadence.
Shape 5 — The org-wide seat (Wayfair, The Access Group, Citi)#
| Deployment | Wayfair: 1,600+ engineers all-in. Access Group: Devin across 250 products, 2,000+ engineers → 5× output. Citi: 40,000-developer rollout, 30% tech-vendor spend cut |
| Mechanism | Not more agents — the seat: every engineer has Devin access, shared playbooks, org-level knowledge |
How to copy: the rollout order in the published cases is consistent — pilot team → measured results → playbook/knowledge standardization → seat expansion. Skipping the knowledge step is why seat rollouts stall: see playbooks guide.
Shape 6 — The process gate (Hamming, Litera, Eight Sleep)#
| Deployment | Hamming: Devin writes 25% of code volume. Litera: regression cycles cut 90%. Eight Sleep: 8× more shipped |
| Mechanism | Devin owns a stage of the pipeline — regression testing, QA — not the whole product |
How to copy: find the pipeline stage with the worst human cost (regression cycles, QA); assign it to a Bot or automation; measure the stage's cycle time before/after.
Anti-patterns the cases imply#
- Tool, not colleague. The high-number deployments give Devin ownership (repos, surfaces, stages), not assignments — same lesson as Guides 1–5.
- Pilot forever. Published results all share one trait: a measured pilot that graduated. Set the graduation criterion upfront (PR share, incidents covered, hours saved).
- No knowledge layer. Every big deployment encodes conventions somewhere — fine-tunes, playbooks, Bot instructions. Without it, quality varies by session.
Verify it#
- [ ] Your rollout doc names one pipeline stage, one metric, and a graduation criterion — before the first session
- [ ] Week-2 checkpoint: can you state Devin's merged-PR share on that stage without querying anyone?
Pitfalls#
- Scaling the pilot that never graduated. Every big number here came after a measured pilot with an exit criterion.
- License-first thinking. Buying seats before naming the owned stage produces shelfware — the pattern is stage → pilot → metric → then seats.
Knowledge check#
- Why do all six deployment shapes share 'an owned stage' as their atom? (Every published number came from a bounded, measurable surface — orgs that deployed 'everywhere' published nothing.)
- What's the graduation criterion and why set it before the pilot? (A metric + threshold that flips pilot → production; without it, pilots drift forever or get killed by vibes.)
- Why does the knowledge layer appear in every large deployment? (Consistency at scale requires shared context — the alternative is N different-quality sessions.)
Sources#
- devin.ai/customers — all cases and numbers cited
- built-with-devin — enterprise index — the registry's verified case table