GitHub COO Kyle Daigle dropped some numbers on X last week that deserve a second look. The platform logged 1 billion commits across all of 2025. It is now seeing 275 million commits per week, a pace that would produce 14 billion for the year if it held steady. Daigle himself doesn't think it will stay linear. He's probably right, but not in the direction most people assume.
GitHub Actions usage tells a similar story: 500 million minutes per week in 2023, 1 billion in 2025, and 2.1 billion as of early April 2026. That's a quadrupling in three years for a service that was already running at enormous scale.
The source of all this new activity isn't a mystery. AI coding agents (Claude Code, OpenAI's Codex CLI, Cursor, and a growing swarm of open-source alternatives) have gone from curiosities to standard workflow components. According to The Information, AI-agent pull requests jumped from roughly 4 million in September to 17 million in March. That's a 4x increase in six months, and each of those PRs triggers CI runs, webhook events, code review bots, and often more agent activity downstream.
The outages nobody can ignore
GitHub has been falling over with uncomfortable regularity. February 2026 alone produced six separate incidents, including a nearly six-hour Actions outage on February 2nd and a multi-hour degradation on February 9th that took down the web UI, pull requests, and Copilot simultaneously. March brought more of the same: Pull Requests went down on the 31st due to a garbage collection misconfiguration, and Actions suffered delays on the 30th from database replication lag. April opened with Copilot agent sessions hitting resource exhaustion on the 1st, followed by rate-limiting that knocked out 40% of Copilot Cloud Agent jobs on the 2nd.
The Register put it bluntly: GitHub is struggling to maintain even three nines of availability. One unofficial reconstruction of their status data suggested uptime dipped below 90% at one point in 2025. For a platform that promises 99.9% in its Enterprise SLA, that's a gap measured in days, not minutes.
Why it keeps breaking
CTO Vladimir Fedorov addressed the situation in a March blog post that was refreshingly specific. The February 9th incident came down to an overloaded database cluster, compounded by a cache TTL reduction they'd made to push a new AI model out faster. Monday traffic surged, client updates spiked read load, and the system tipped over. Responders lacked the granular switches needed to shed load selectively.
Lori Hochstein, a software engineer who analyzed the CTO's post, noted something important: databases are hard to scale horizontally, and saturation events from unexpected load are exactly the kind of failure you'd expect when traffic doubles in a few months. The mitigation playbook for stateless services doesn't apply.
And then there's the Azure migration. GitHub began moving its core platform off its Virginia data centers and onto Microsoft Azure in October 2025. Actions and Copilot migrated in 2024. Pages and Packages went in 2025. But the core platform and databases are still mid-migration, and running split traffic between two environments is, as one Hacker News commenter put it, a recipe for "constant background cognitive load and surface area for bugs."
The agent economics problem
Here's what I keep coming back to: agents don't pay. A human developer on a free GitHub account generates some commits, maybe a few CI runs. An AI agent on the same account can generate hundreds of commits, dozens of PRs, and thousands of Actions minutes in a single afternoon. The infrastructure cost per "user" has fundamentally changed, but the pricing model hasn't caught up.
GitHub's Octoverse 2025 report celebrated nearly 1 billion commits and 36 million new developers. But the 2026 numbers Daigle shared aren't being driven by 14x more developers showing up. They're being driven by agents that treat GitHub's API like a utility, which it basically is, except utilities charge for consumption.
GitGuardian's 2026 Secrets Sprawl report offers an indirect measure of the scale shift: public GitHub commits grew 43% year over year to 1.94 billion in 2025, with 54% of active developers having made their first commit that year. The platform isn't just getting busier. Its user base is skewing toward less experienced developers amplified by AI tools, and the security implications are measurable. AI-assisted commits leaked credentials at roughly double the baseline rate.
Can GitHub actually handle this?
Fedorov's blog post listed the near-term fixes: redesign the user cache system, audit critical infrastructure capacity, isolate key dependencies so Actions and Git don't share failure domains with everything else. Standard reliability engineering. The kind of work that should have been done before the traffic arrived, but wasn't.
A startup called Pierre Computer, founded by Bootstrap creator Jacob Thornton, claims to have built an agent-native Git host that handled 15,000 new repos per minute in a sustained peak. GitHub was doing about 230 per minute. Gergely Orosz noted in The Pragmatic Engineer that Mitchell Hashimoto, the Ghostty founder, went so far as to suggest GitHub should buy Pierre, shut down Copilot, and refocus entirely on being infrastructure for agentic workflows.
That won't happen, obviously. Copilot is Microsoft's AI revenue story. But the frustration is real. Zig already left GitHub over what its maintainers described as Microsoft's AI obsession degrading the core service. When your most technically sophisticated users start looking for exits, you have a problem that blog posts can't fix.
What comes next
The Copilot training data announcement didn't help the mood. From April 24, GitHub will use interaction data from Copilot Free, Pro, and Pro+ users to train its AI models unless they opt out. Combine that with the outages, the Azure migration instability, and the agent traffic explosion, and you get a platform that's simultaneously trying to be the world's code host, the world's CI/CD provider, and the world's AI coding platform, while struggling to keep the lights on for any of them.
Rate limits for agents feel inevitable. So do agent-specific pricing tiers. The question is whether GitHub implements them proactively or waits until the next multi-hour outage forces it. Given the pattern so far, I'd bet on the latter.
