GitHub has started sending some traffic through Amazon Web Services after AI coding agents pushed its systems past the limits it had prepared for. The move highlights a sudden shift in how code platforms are used, with automated tools running tasks around the clock rather than during normal working hours.
Rapid growth in AI-driven activity
Usage of GitHub Actions, the service that runs automated checks and builds, rose from 500 million minutes across 2023 to more than 2 billion minutes in a single week this year. Pull requests opened by AI agents increased from roughly 4 million last September to over 17 million by March. Weekly commits reached 275 million, putting the platform on course for 14 billion this year—fourteen times the previous total.
GitHub had begun scaling capacity for a tenfold increase last October, only to revise that target upward to thirty times by February. Even so, the company recorded nine service incidents in one month, with availability dropping below 99 percent. Microsoft, which owns GitHub, quietly added Amazon infrastructure to keep services stable.
Why human assumptions no longer hold
Traditional capacity planning assumed developers work in patterns shaped by sleep, weekends and time spent reviewing changes. AI agents follow no such rhythm: they can open a request, trigger checks, review results and repeat without pause. This removes the natural limits that once kept overall demand predictable.
For small businesses that rely on GitHub to host code, run tests or manage updates, the change brings practical risks. Automated tools may now trigger far more activity than expected, raising the chance of slower response times or added costs if usage-based pricing applies. Firms that have started using AI assistants for routine coding tasks could see their own workflows affected when the underlying platform struggles.
What to watch next
Platform operators are likely to redesign limits and pricing around nonstop agent activity rather than human schedules. Business owners should check whether their current tools assume human-paced use and consider what would happen if thousands of automated processes ran simultaneously. Those relying on GitHub Actions or similar services may want to review usage reports and test fallback options before demand grows further.