There is a consistent pattern in our Awario sweeps of Devin coverage: the most operationally useful writing is in Japanese. While English-language coverage argues about whether agents replace engineers, the Zenn/Qiita/note cluster publishes migration guides, YAML, and honest failure notes. This week’s sweep surfaced enough of them to assemble into something like a playbook — with each claim below attributed to the team that reported it.
The baseline case study: Devin joins the Zenn team
The canonical entry is Zenn’s own tech blog, where dyoshikawa documented a production trial on Zenn’s real, closed codebase — chosen deliberately, because greenfield demos prove nothing. The reported outcome: about 30 merged Devin PRs in 10 business days, against a personal baseline of one to two PRs per day, and “70–80%+” of one new feature implemented by Devin. The same post is candid about the failure mode: handing over a requirement wholesale didn’t work; implementation-level instructions did.
Their operating rules translate cleanly:
Rule 1
Split tasks aggressively. Small scope means the agent escapes the "pit of death" of repeated failed fixes — and small PRs are mergeable PRs.
Rule 2
Front-load context: the why, the approach, target file paths, reference files. Put reusable context in Knowledge so it's DRY across sessions.
Rule 3
Tests are load-bearing. Without a test suite the agent can't run its own feedback loop — Zenn's Devin usage skews backend partly because frontend tests were thin.
They also name Devin-specific advantages that editor-embedded agents don’t share: sandboxed cloud environments make command execution low-anxiety, and per-task isolated environments let stateful backend work (think conflicting DB migrations) run in parallel without branch-juggling.
The budget question: what to do with leftover ACUs
The most relatable post in the sweep is kyonsi’s six-month retrospective on a team that kept ending months with unspent ACUs. Their three keepers:
- DeepWiki as the team’s question desk. Not the generated wiki itself — the ability to interrogate a codebase (“where is this called from?”, “does this work in a Server Component?”) with answers linked to source. They extended it beyond code: meeting notes auto-exported from Google Meet, pushed to a repo, indexed by DeepWiki, so past architecture decisions became queryable.
- CI code review with session reuse. They wired Devin into PR review via the API following Cognition’s own Devin 101 guide — then improved on it, tagging sessions per PR and reusing them across pushed commits instead of paying session startup and ACU cost each time. Their workflow YAML is in the post. They flag the trade-off themselves: reused sessions accumulate context, and they’re still validating whether accuracy degrades.
- Prompts as Playbooks, not workflow strings — moving review prompts out of GitHub Actions and into maintained playbooks.
Two shorter reports complete the picture. Photoruction points Devin Review at the Dependabot queue: encode merge policy in a REVIEW.md that Devin Review reads, and the dependency-bump pile nobody wants to review starts moving. And PKSHA describes a designer running a design refactor entirely from Slack — Devin reading code, running tests, and opening PRs while the designer never opened an editor. The notable part isn’t autonomy; it’s who got access to changing the product.
Method: All productivity figures (30 PRs / 10 days, 70–80% of a feature, ACU practices) are self-reported by the teams in the linked posts — field reports, not Cognition telemetry. The Zenn team trial was published March 2025 and updated since; the ACU retrospective is from December 2025, updated July 2026. Product behavior is cross-checked against docs.devin.ai (fetched 2026-08-12).
References
- ZennチームにもDevinがジョインしました — production trial: 30 PRs / 10 business days, task-splitting, Knowledge, tests, sandbox parallelism.
- DevinのACUを余らせていた私たちが、半年かけて辿り着いた3つの活用法 — DeepWiki as question desk, session-reuse review CI, playbook prompts.
- Devinを使ってDependabotのマージハードルを下げる — REVIEW.md policy for dependency PRs.
- Devinを使ってデザイナーがSlackだけでデザインリファクタリングした話 — designer-driven refactor via Slack.
- Devin 101: Automatic PR Reviews with the Devin API — the official how-to the CI pattern extends.
- DeepWiki · Devin Review · Knowledge — product docs.
Editorial from Devin Central — a fan news desk, not Cognition. Devin is a trademark of Cognition.