Claude Code, Devin reshape coding as AI agents take over

The article frames AI coding agents as a distinct step beyond autocomplete: rather than predicting the next line, an agent keeps a work plan, edits files, runs tests and revises until the task passes. It sorts the tools into three styles — terminal agents like Claude Code that work across a whole repository, IDE agents like Cursor that show diffs for approval, and cloud agents like Devin that take a task and return a pull request. Rule files such as CLAUDE.md and AGENTS.md carry project conventions between sessions, and permission settings decide how much the agent may do without asking.

The evidence the article offers is mostly company case studies. Rakuten had Claude Code work for 7 hours on vLLM, a 12.5-million-line open-source library, and reported cutting lead time from 24 business days to 5. Nubank used Devin on a migration of more than 6 million lines, reporting that work once forecast at 18 months and over 1,000 engineers finished in weeks at more than 20x lower cost. Classmethod reported up to 10x productivity and an 80% cut in code-review time, and ULS Group used Devin on about 1.5 million test steps, cutting effort from 200 person-months to 50. Set against this, Veracode's 2025 survey found 45% of code samples from over 100 AI models failed security tests, and 46% of developers told Stack Overflow they do not trust AI tool output.

That contrast is the crux. The outcome likely hinges on process rather than tool choice — clear task boundaries, limited permissions, rule files and automated test and CI checks. Teams with those in place may be able to widen what they delegate; teams without them risk carrying generated errors into production. The article's own answer to the skeptics is to start small, on tasks where success or failure is easy to judge.