GitHub goes down for 8 hours mid-workday: agent traffic overwhelms its infrastructure
On Monday afternoon, GitHub was down for 8 hours. A bug in a load balancer sidecar saturated the servers, and requests started to fail. Then the domino effect arrived: Visual Studio Code retried each failed login up to 10 times in a row. That retry storm multiplied the requests and took the servers down completely.
The classic solution for this pattern is exponential backoff: you retry, but you double the wait after each failure. One second, then two, then four. That way you do not hammer servers that are already saturated.
The underlying problem is scale. In all of 2025, GitHub received about 1 billion commits. In 2026 it runs at 275 million per week: every 4 weeks it matches the traffic of the entire previous year. AI agents generate most of that increase.
The alternative status page shows the real picture: 93.73% uptime over 90 days, one and a half hours down every day. In February they hit their low point at 84.31%. GitHub now migrates to Azure and prepares its infrastructure to multiply its current scale by 10 to 30. Watch in the video.
All the news we covered