Continuous AI in software engineering

Goedecke compares AI to tests and type checks. Nobody decides to run those manually each time; they run continuously through CI, pre-push hooks and IDE highlighting, and a developer benefits from them even without consciously triggering them. He argues AI in the software development lifecycle should work the same way, a pattern he calls “continuous AI,” a term he attributes to GitHub. Having Copilot PR reviews turned on by default changed his mind. One in five or ten catches something he missed; skimming past the rest costs him almost nothing.

Fichamento

#

• Draws the core analogy explicitly: using AI “on tap” (going to get it when there’s a problem) is like manually running unit tests or a type checker instead of having CI and IDE highlighting do it continuously.
• Defines continuous AI as any way AI tooling gets automatically integrated into existing development flows, running via CI, hooks or a schedule on its own. Tools like Claude Code or Devin still need a developer to actively invoke them each time.
• Examples given: automated AI-driven PR reviews, AI-driven issue/PR labeling, automated daily or weekly summary rollups, and Copilot/Cursor-style autocomplete.
• Copilot PR reviews turned on by default is what converted him: most individual reviews don’t add value, but one in five or ten catches something real, and skimming four low-value reviews to reach the useful one is a cheap trade.
• He extends the same logic to internal reporting: weekly rollups that summarize project status for the management chain are, in his view, safe to delegate to AI most of the time, reserving human attention for when a project is actually going badly.
• Describes adding the GitHub Models permission to the native GitHub Actions access token as one of his most satisfying small wins: free inference plus Actions’ triggers (PR/issue events, pushes, cron) makes continuous AI cheap to experiment with, without anyone needing to approve a subscription first.
• His closing position: even in a future where agentic models handle most or all coding, he expects layers of automated checks and organizational AI tasks to persist as ambient infrastructure regardless.