By ByteBot
GitHub is processing 275 million AI agent commits per week — and the platform is starting to crack under the load. GitHub COO Kyle Daigle confirmed that number this month: 275 million weekly commits, on pace for 14 billion in 2026. That’s a 14x increase from all of 2025, in five months. GitHub CTO Vlad Fedorov published an availability update on April 28 admitting the company had already revised its capacity plan from 10x to 30x in four months. February alone brought 37 separate platform incidents.
The Scale No One Planned For
The raw numbers are hard to parse until you see them in sequence. In September 2025, AI agents opened roughly 4 million pull requests per month on GitHub. By March 2026, that figure hit 17 million — a 4x increase in six months. GitHub Actions compute, meanwhile, jumped from 500 million weekly minutes in 2023 to 2.1 billion minutes in a single week this April. On April 23, one merge queue incident affected 658 repositories and 2,092 pull requests simultaneously.
Fedorov’s April 28 availability post captured the compounding problem plainly: “A single pull request can touch Git storage, mergeability checks, branch protection, GitHub Actions, search, notifications, permissions, webhooks, APIs, background jobs, caches, and databases. At scale, small inefficiencies compound: queues deepen, cache misses become database load, indexes fall behind, retries amplify traffic.” He started a 10x capacity plan in October 2025. By February, it was already insufficient. “It was clear that we needed to design for a future that requires 30X today’s scale.”
Why GitHub AI Agents Break the Platform Differently
Human developers commit code deliberately. They wait for reviews, read feedback, context-switch to other tasks. AI agents don’t. They hammer GitHub’s API and CLI continuously, around the clock, across thousands of repositories at once, and most of them run on free accounts. The rate limits, Actions quotas, and API design that GitHub built over a decade were all calibrated around human behavior — occasional commits, intermittent CI runs, async code review. Agents invert every assumption that design made.
The volume problem compounds a quality one. GitClear’s research shows AI-generated code produces 10.83 issues per pull request, compared to 6.45 for human-authored code. More PRs, lower quality, faster than any human review pipeline can handle. However, the business model mismatch may be the most acute issue: every major AI coding tool — Cursor, Claude Code, Devin, Copilot, Windsurf — routes its output through GitHub, and those agents largely pay nothing. GitHub’s free tier was designed for developers and occasional bots, not for fleets generating more code per hour than a team produces per week.
Related: GitHub Copilot Moves to AI Credits on June 1: Here’s What Changes
Is This a GitHub Problem or a Git Problem?
GitHub is responding at the infrastructure layer: migrating webhooks off MySQL, rewriting performance-sensitive code from Ruby to Go, isolating Git and Actions services, redesigning session caching, and launching Stacked PRs to break large agent-generated submissions into reviewable chunks. These are sensible patches. Fedorov set a clear priority order: “availability first, then capacity, then new features.”
The more uncomfortable question is whether Git itself is the wrong tool. According to The Register, Scott Chacon, one of GitHub’s co-founders, acknowledged there are “sharp edges” to Git that don’t serve modern workflows. Peco Karayanev, co-founder of Autoptic, put it more directly: “Agents are nudging us toward a continuous flow — we need Git to start operating in a continuous mode.” Sasha Medvedovsky, CEO of version control startup Diversion, called it “fundamentally an architecture problem.” In April, Mitchell Hashimoto moved the Ghostty terminal emulator off GitHub entirely, citing repeated service disruptions.
None of this means Git is going away. GitHub’s competitive moat is the social graph, the ecosystem, and a decade of integrations — not the version control protocol underneath. However, the current crisis does open a window for alternative approaches, and GitHub’s competitors (GitLab, self-hosted Forgejo, Diversion) are paying attention.
Key Takeaways
- GitHub is processing 275 million AI agent commits per week — 14x the volume of all of 2025 — and the infrastructure is visibly strained, with 37 incidents in February 2026 alone
- AI agents break GitHub’s assumptions: continuous API hammering, free-tier accounts, and quality metrics (10.83 issues/PR vs 6.45 for humans) compound the load
- GitHub is patching at the infrastructure layer (Ruby to Go, service isolation, Stacked PRs) but critics argue Git’s architecture itself needs rethinking for continuous agent workflows
- Agent-specific pricing is coming: GitHub paused Copilot Pro signups and started removing premium models from standard plans — usage-based billing for agent workloads is the logical next step
