The overnight shift
Developers are running autonomous coding sessions while they sleep using Ralph Wiggum, an official Anthropic Claude Code plugin that loops tasks until completion. The approach inverts traditional AI pair programming: write requirements at 5pm, review finished code at 9am.
The mechanism is straightforward. A bash loop feeds the same prompt back to Claude repeatedly. Each iteration sees prior code, error logs, and git history. The loop continues until acceptance criteria are met. One developer's first overnight session cost $12 in API calls and delivered 47 commits on a deferred refactoring task.
What actually works
Three weeks of field testing shows clear patterns. Good candidates: refactoring with defined goals, feature implementation from detailed specs, test coverage expansion, migration tasks. Poor candidates: exploratory work, architectural decisions, anything requiring human judgment.
The quality of autonomous output correlates directly with requirement clarity. "Add user authentication" produces constant clarification requests. A spec defining JWT implementation, endpoint requirements, bcrypt hashing, and 80% test coverage runs autonomously to completion.
Professionals have converged on a three-phase methodology: detailed requirements files with acceptance criteria, AI-generated implementation plans for human review, then autonomous building with test-driven self-correction.
Production deployments
This isn't toy work. YC hackathon teams shipped six-plus repos overnight for $297 in API costs. One developer completed a $50,000 contract for under $300. Geoffrey Huntley ran a three-month loop that built an entire programming language.
The catch: Claude API enforces five-hour session limits. Ralph handles resets via 60-minute waits, making overnight and multi-day loops viable. Community forks add production wrappers like token tracking and circuit breakers, addressing operational gaps in the core plugin.
The productivity gap
Early adopters report 5-20x productivity improvements on suitable tasks. The gap between developers using autonomous loops and those who aren't is widening. Side projects that required months can be validated in days. Contract work that wasn't profitable becomes worth taking.
The skill becomes knowing what to automate and how to specify it clearly. Less time on repetitive implementation, more time on architecture, requirements, and review. That's the shift professional developers are processing now.