What Happened
On June 16, Microsoft confirmed it is routing GitHub traffic through Amazon Web Services after AI coding agents overwhelmed the platform's own infrastructure. AI agent-opened pull requests grew from 4 million per month in September 2025 to 17 million by March 2026 — a 325% increase in six months. Claude Code alone went from roughly 100,000 commits per week to 2.6 million, a 25x jump in the same period. GitHub Actions weekly compute hit 2.1 billion minutes in early 2026, up from 500 million for all of 2023. The platform logged nine service incidents in May and ten in April, with availability dropping to roughly 88.4% in June — well below the 99.9% enterprise SLA. Microsoft is now relying on its largest cloud competitor to keep its most strategically important developer platform running.
The Architecture Problem Behind the Numbers
GitHub's infrastructure was designed for human-scale usage. Its core is a Ruby on Rails monolith with tightly coupled services. When one component fails — an authentication database under load — it cascades simultaneously into Actions outages, Copilot failures, and web UI degradation. That coupling wasn't a problem when traffic grew at human rates. Agent traffic doesn't grow at human rates. Microsoft's infrastructure team planned for 10x capacity expansion in October 2025 and realized they needed 30x by February 2026. Architecture decisions that were invisible at human scale became critical failure modes at agent scale.
Why Agent Traffic Is Different
Human developers are rate-limited by biology. An agent has no natural throttle. A single AI coding session can open dozens of pull requests, trigger CI pipelines, iterate on test failures, and push fixes in the time a human developer reads their morning messages. When millions of developers deploy agents simultaneously, the platform sees a traffic profile that looks nothing like the usage patterns it was architected for. This isn't just a capacity problem — it's a behavioral mismatch. The platform was designed around human pacing, and agents don't pace.
What This Means for Platform Builders
Anyone building a platform that AI agents will use at scale — APIs, code repositories, financial data feeds, document processing pipelines, compliance tooling — needs to re-examine capacity assumptions now, not when the crisis hits. The GitHub situation previews what happens when you plan for human-scale growth in a world where agents compound usage exponentially. Specific design questions: Does your authentication layer hold when a single customer deploys 500 parallel agents? Does a single component failure cascade into your entire platform? Do your rate-limiting semantics make sense when the caller isn't a person?
The Broader Signal
GitHub routing through AWS is a case study worth tracking closely. The underlying trend — AI agents becoming the dominant source of platform traffic — won't reverse. GitHub will resolve the immediate crisis through capacity expansion and likely an architectural evolution away from its monolith. The durable question is which platforms are being designed from the start for agent-scale usage, and which will face their own version of this crisis when their traffic profile shifts in ways the original architecture can't handle. The teams building infrastructure today that treat agent-generated load as a first-class design constraint will be in a structurally different position than those treating it as an edge case.